The Oxford Handbook of Inflection 9780199591428, 0199591423

Citation preview

the oxford ha ndb o o k of

IN F L EC TI O N

OXFORD HANDBOOKS IN LINGUISTICS Recently published

THE OXFORD HANDBOOK OF ARABIC LINGUISTICS Edited by Jonathan Owens

THE OXFORD HANDBOOK OF COMPOSITIONALITY Edited by Markus Werning, Wolfram Hinzen, and Edouard Machery

THE OXFORD HANDBOOK OF COMPOUNDING Edited by Rochelle Lieber and Pavol Štekauer

THE OXFORD HANDBOOK OF CONSTRUCTION GRAMMAR Edited by Thomas Hoffman and Graeme Trousdale

THE OXFORD HANDBOOK OF CORPUS PHONOLOGY Edited by Jacques Durand, Ulrike Gut, and Gjert Kristoffersen

THE OXFORD HANDBOOK OF DERIVATIONAL MORPHOLOGY Edited by Rochelle Lieber and Pavol Štekauer

THE OXFORD HANDBOOK OF GRAMMATICALIZATION Edited by Heiko Narrog and Bernd Heine

THE OXFORD HANDBOOK OF HISTORICAL PHONOLOGY Edited by Patrick Honeybone and Joseph Salmons

THE OXFORD HANDBOOK OF THE HISTORY OF ENGLISH Edited by Terttu Nevalainen and Elizabeth Closs Traugott

THE OXFORD HANDBOOK OF THE HISTORY OF LINGUISTICS Edited by Keith Allan

THE OXFORD HANDBOOK OF LABORATORY PHONOLOGY Edited by Abigail C. Cohn, Cécile Fougeron, and Marie Huffman

THE OXFORD HANDBOOK OF LANGUAGE AND LAW Edited by Peter Tiersma and Lawrence M. Solan

THE OXFORD HANDBOOK OF LANGUAGE EVOLUTION Edited by Maggie Tallerman and Kathleen Gibson

THE OXFORD HANDBOOK OF LINGUISTIC ANALYSIS Second Edition Edited by Bernd Heine and Heiko Narrog

THE OXFORD HANDBOOK OF LINGUISTIC FIELDWORK Edited by Nicholas Thieberger

THE OXFORD HANDBOOK OF SOCIOLINGUISTICS Edited by Robert Bayley, Richard Cameron, and Ceil Lucas

THE OXFORD HANDBOOK OF TENSE AND ASPECT Edited by Robert I. Binnick

THE OXFORD HANDBOOK OF THE WORD Edited by John R. Taylor For a complete list of Oxford Handbooks in Linguistics, please see pp. 689–690

the oxford handbook of ......................................................................................................

I NF LECTI O N ......................................................................................................

Edited by

MAT THEW BA ERM A N

1

3 Great Clarendon Street, Oxford, ox2 6dp, United Kingdom Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide. Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries © editorial matter and organization Matthew Baerman 2015 © the chapters their several authors 2015 The moral rights of the authors have been asserted First Edition published in 2015 Impression: 1 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, by licence or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this work in any other form and you must impose this same condition on any acquirer Published in the United States of America by Oxford University Press 198 Madison Avenue, New York, NY 10016, United States of America British Library Cataloguing in Publication Data Data available Library of Congress Control Number: 2015933904 ISBN 978–0–19–959142–8 Printed and bound by CPI Group (UK) Ltd, Croydon, CR0 4YY Links to third party websites are provided by Oxford in good faith and for information only. Oxford disclaims any responsibility for the materials contained in any third party website referenced in this work.

Contents

...........................

List of Abbreviations List of Contributors

ix xvii

1. Introduction Matthew Baerman

1

PART I BUILDING BLOCKS 2. The Morpheme: Its Nature and Use Stephen R. Anderson

11

3. Features in Inflection Greville G. Corbett

35

4. Inflectional Exponence Jochen Trommer and Eva Zimmermann

47

PART II PARADIGMS AND THEIR VARIANTS 5. Inflectional Paradigms James P. Blevins

87

6. Inflection Classes Gregory Stump

113

7. Paradigmatic Deviations Matthew Baerman

141

8. Phonology Gunnar Ólafur Hansson

161

9. Periphrasis and Inflection Andrew Spencer and Gergana Popova

197

vi

contents

PART III CHANGE 10. Diachrony Claire Bowern

233

11. Contact-Induced Change Maarten Kossmann

251

PART IV COMPUTATION 12. Modelling Inflectional Structure Dunstan Brown

275

13. Machine Learning of Inflection Katya Pertsova

297

14. Machine Translation Ondrej ˇ Bojar

323

PART V PSYCHOLINGUISTICS 15. Inflectional Morphology in Language Acquisition Sabine Stoll

351

16. Disorders Matthew Walenski

375

PART VI SKETCHES OF INDIVIDUAL SYSTEMS 17. Verbal Inflection in Iha: A Multiplicity of Alignments Mark Donohue

405

18. Inflection in Pulaar Fiona Mc Laughlin

419

19. Lithuanian Inflection Axel Holvoet

447

20. Chamorro Inflection Thomas Stolz

465

contents

vii

21. Inflection in Murrinh-Patha Rachel Nordlinger

491

22. Aymara Inflection Matt Coler

521

23. Inflection in Nen Nicholas Evans

543

24. Stem-internal and Affixal Morphology in Shilluk Bert Remijsen, Cynthia L. Miller-Naud´e, and Leoma G. Gilley

577

References Author Index Language Index Subject Index

597 669 683 687

List of Abbreviations .............................................................

α, β, γ

undergoer prefixes, underspecified for TAM value

>

acting upon

|

(disjunction/ between morphosyntactic values), e.g. 2|3sg ‘2nd or 3rd singular’

abl

ablative

abs

absolutive

Accom

accompaniment

acc

accusative

ad

additive

adess

adessive

ag

agentive nominalizer

Agr

agreement

agt

agentive/instrumental

all

allative

anim

animate

anmz

action nominalizer

ant

anticipatory

antip

anticipatory

appl

applicative

ARG-ST

argument sructure

asal

argument salience

ass

assentive

assert

assertive

ATR

advanced tongue root

att

attributive

aux

auxilary

awa

away

BA

Brodmann’s areas

ben

benfactive

x

list of abbreviations

bfr

buffer

CAT

computer-assisted translation

caus

causative

cc

concentrative

cf

counterfactual

cfy

confirmatory

clf

classifier

clf:anim

noun classifier animates

clf:human

noun classifier humans

clf:thing

noun classifier neuter/residue class

clf:time

noun classifier times and places

clf:veg

noun classifier non-meat food

cm

class marker

cnj

conjectural

cnj2

conjectural phrase-final

coll

collective

com

comitative

cond

conditional

conj

conjugation

cont

continuous

cop.vbz

copulative verbalizer

cp

comparative

cpl

completive

CSL

Center for Speech and Language

cur

current

cvb

converb

dat

dative

decl

declarative

def

definite

dem

demonstrative

df

diffuse

dist

distal

dl

delimitative

dm

discourse marker

DNF

disjunctive normal form

list of abbreviations DP

Declarative/Procedural Model

du

dual

dw

downward

emph

emphatic marker

erg

ergative

evid

evidential

exc

exclamatory

excl

exclusive

exist

existential

f

feminine

foc

focus

FSA

finite state automaton

FSM

finite state morphology

FST

finite state transducers

fugal

centrifugal spatial deixis

fut

future

fut:irr

future irrealis

GB

Government and Binding

gen

genitive

GET

generative-enumerative syntax

GNA

Generalized Nonlinear Affixation

gofoc

goal focus

H

high tone

HD

Huntingdon’s disease

HPSG

Head-driven Phrase Structure Grammar

HSE

Herpes Symplex Encephalitis

icpl

incompletive

IE

Indo-European

imp

imperative

ina

interactive

incl

inclusive

ind

indicative

indef

indefinite

xi

xii

list of abbreviations

iness

inessive

inf

infinitive

infr

inferential

InO

indirect object

ins

instrumental

intens

intensive

intr

intransitive

IO

input–output

ipfv

imperfective

ir

interrogative

irr

irrealis

iter

iterative

iw

inward

L

low tone

LID

lexical index

lim

limitative

link

linker particle

LKŽ

Lietuvių kalbos žodynas [Lithuanian Dictionary]

LM

language model

loc

locative

loc.vbz

verbalizer

m

masculine

mo

momentaneous

MDL

Minimum Description Length

MLE

maximum likelihood estimate

mlt

multiplier

mod

modal verb

MRL

morphologically rich language

MS

Morphological Structure

MT

machine translation

MTS

model-theoretic syntax

n

neuter

NCA

non-concatenative allomorphy

nd

non-dual

neg

negative

nevid

non-evidential

list of abbreviations nfut

non-future tense

nmlz

nominalizer

nom

nominative

nsg

non-singular

nsib

non-sibling

num

numeral

O/obj

object

obl

oblique

OCR

optical character recognition

PAC

Probably Approximately Correct

pass

passive

pat

patientive

PBMT

phrase-based machine translation

pc

paucal

pdr

perdurative

PEP

Paradigm Economy Principle

perl

perlative

perm

permanent property

PET

positron emission tomography

petal

centripetal spatial deixis

PFM

Paradigm Function Morphology

PHON

phonology

p:ipfv

past imperfective

pfv

perfective

pl

plural

PM

Prosodic Morphology

PN

proper noun

pn

pronominal

PNG

Papua New Guinea

poss

possessor/possessive

pot

potential

pret

preterite

prf

perfect

pri

primordial

xiii

xiv

list of abbreviations

priv

privative

prog

progressive

proj.imp

projected imperative

pron

pronoun

prox

proximal

prp

propogative

prs

present

pst

past

pst:irr

past irrealis

pt

perlative

ptcp

particple

ptnmlz

patient nominalization

pvb

preverb

rcand

reduplicand

rdp

reduplication

r.e.

recursively enumerable

re

resultative

rec

recent past

recip

reciprocal

ref

referential

refl

reflexive

rem

remote past

repet

repetitive action

RM

realize morpheme

RNC

Russian National Corpus

rr

reflexive/reciprocal

S/sbj

subject

sbjv

subjunctive

sfx

suffix

sg

singular

semel

semelfactive

sib

sibling

sim

simple tense

SLI

Specific Language Impariment

list of abbreviations smlf

semelfactive

SMT

statistical machine translation

sou

source

spat

spatial

spec

specifier

ss

same subject simultaneous

str

sustainer

subr

subordinator

TAF

transderivational antifaithfulness

TAM

tense, aspect, mood

TETU

The Emergence of the Unmarked

tf

transformative

top

topic/topicalizer

tow

toward

tpst

today’s past

tr

transitive

trans

transitivizer

tube

tubular classificatory marker

u

undergoer

UG

Universal Grammar

voc

vocative

VoiObs

voiced obstruent

vsal

verb salience

WP

Word and Paradigm

ypst

yesterday’s past

xv

List of Contributors ...........................................................

Stephen R. Anderson is the Dorothy R. Diebold Professor of Linguistics at Yale University. His academic interests in linguistics include phonology, morphology, the theory of clitics, the history of the discipline, and the descriptive study of a range of languages. His current work is concerned with the structure of Rumantsch and with the biology of language, including animal communication and the evolution of the language faculty. His most recent book is Languages: A Very Short Introduction (OUP, 2012). Matthew Baerman is a research fellow in the Surrey Morphology Group at the University of Surrey. His research focuses on the typology, diachrony, and formal analysis of inflectional systems, with a particular concentration on phenomena whose interpretation is problematic or controversial. James P. Blevins is Reader in Morphology and Syntax in the University of Cambridge. His current research is concerned mainly with the structure and complexity of inflectional systems. These issues are approached from the standpoint of contemporary word and paradigm models, interpreted from a largely information-theoretic and discriminative perspective. Ondˇrej Bojar is an assistant professor at the Institute of Formal and Applied Linguistics (ÚFAL) at the Faculty of Mathematics and Physics, Charles University in Prague. After graduating there in computational linguistics, Ondˇrej has been working on syntactic analysis and lexical acquisition. A visit to Hermann Ney’s department at RWTH Aachen University caused a major twist in Ondˇrej’s interests and, since 2006, most of his research revolves around machine translation, targetting mainly morphologically rich languages. Claire Bowern is Associate Professor of Linguistics at Yale University. Her 2004 PhD is from Harvard University and examined the historical morphology of complex verb constructions in a family of non-Pama-Nyungan (Australian) languages. Her research focuses on the Indigenous languages of Australia, and is concerned with documentation/description and prehistory. She is the author (with Terry Crowley) of An Introduction to Historical Linguistics (OUP, 4th edn), and editor of Australian Languages: Classification and the Comparative Method (with Harold Koch) and Morphology and Language History (with Bethwyn Evans and Luisa Miceli). She serves on the editorial boards of Diachronica and Journal of Historical Linguistics and is co-editor (with Ashwini Deo) of the historical section of Language and Linguistics Compass.

xviii

list of contributors

Dunstan Brown holds an Anniversary Chair in the Department of Language and Linguistic Science, University of York. His research interests include autonomous morphology, morphology–syntax interaction, and typology. Much of his work focuses on understanding morphological complexity, such as syncretism (The Syntax–Morphology Interface: A Study of Syncretism, with Matthew Baerman and Greville G. Corbett, CUP 2005), as well as computational modelling of morphological systems (Network Morphology, with Andrew Hippisley, CUP 2012). Matt Coler is Group Leader and Scientific Researcher in the Cognitive Systems Group at INCAS3 . Having described a previously-undocumented variety of Peruvian Aymara during the course of his doctoral appointment, Matt not only continues research into less familiar Aymara varieties, but also focuses on the implementation of language and knowledge into technological ‘cognitive’ devices. Greville G. Corbett is Distinguished Professor of Linguistics at the University of Surrey, where he leads the Surrey Morphology Group. He works on the typology of features, as in Gender (1991), Number (2000), Agreement (2006), and Features (2012), all published by Cambridge University Press. His recent research has been within the canonical approach to typology. He is one of the originators of Network Morphology, and publications on morphology include The Syntax–Morphology Interface: A Study of Syncretism (with Matthew Baerman and Dunstan Brown, CUP 2005). Mark Donohue is a linguist at The Australian National University, interested in most aspects of language structure, and very interested in developing models to help us unravel more detailed linguistic histories of languages seen as separate structural units. He began working on languages from New Guinea and Indonesia, and has more recently started work with language and language contact situations in the Himalayas. Nicholas Evans is Professor of Linguistics at the Australian National University. His research focuses on Australian and Papuan languages, typology, and anthropological linguistics. His most important works are grammars of Kayardild (1995) and Bininj Gun-wok (2010), Dying Words: Endangered Languages and What They Have to Tell Us (2010) and (with Stephen Levinson) The Myth of Language Universals (2009). Leoma G. Gilley is the director for training for SIL in Africa. She oversees training in translation and language development at BA through PhD level in partnership with various academic institutions across Africa. Previously, she was an assistant professor at the University of Khartoum. Her main research interests are phonology, Nilo-Saharan languages, and syntax. Leoma has worked with the Shilluk people on their language since 1983.

list of contributors

xix

Gunnar Ólafur Hansson is Associate Professor of Linguistics at the University of British Columbia. His research focuses on phonological theory, the interfaces of phonology with phonetics and morphology, and the interplay of diachronic and synchronic factors in shaping phonological systems and cross-linguistic typologies. A particular area of interest is non-adjacent dependencies in segmental phonology, such as consonant harmony. Axel Holvoet is professor of Baltic linguistics at the University of Warsaw. His research interests include Baltic and Slavonic syntax and morphosyntax, both from a historical and a synchronic point of view. He has published on case and grammatical relations as well as on mood, modality, and evidentiality in Baltic. Maarten Kossmann (PhD Leiden 1994) currently teaches General and African Linguistics at Leiden University. He has worked extensively on the description, comparison, and stylistics of Berber languages in Northern Africa, as well as on questions of language contact in general and in the circumsaharan zone in particular. Fiona Mc Laughlin is an Associate Professor of Linguistics and African Languages and Chair of the Department of Linguistics at the University of Florida. Her research areas include the phonology and morphology of the Atlantic languages, language contact in West Africa, and the sociolinguistics of African cities. She has taught at universities in Senegal and Niger, and is a former director of the West African Research Center in Dakar. Cynthia L. Miller-Naudé is senior professor in the Department of Hebrew at the University of the Free State (Bloemfontein, South Africa). She was trained in the Department of Linguistics and the Department of Near Eastern Languages and Civilizations at the University of Chicago. Cynthia’s main interests are syntax and pragmatics; much of her research involves the ancient Northwest Semitic languages. Since 1992 she has been involved in researching the morphology and syntax of Shilluk. Rachel Nordlinger is Associate Professor of Linguistics at the University of Melbourne, Australia. Her research focuses on the description and documentation of Australia’s indigenous languages, including Bilinarra, Wambaya, and Murrinh-Patha. Rachel has also published on syntactic and morphological theory, and in particular the challenges posed by the complex grammatical structures of Australian Aboriginal languages. She is the author of numerous academic articles in international journals, and four books, including A Grammar of Wambaya (Pacific Linguistics 1998) and Constructive Case: Evidence from Australian languages (CSLI Publications 1998); and A Grammar of Bilinarra (with Felicity Meakins, Mouton de Gruyter 2014).

xx

list of contributors

Katya Pertsova is Assistant Professor in Linguistics at the University of North Carolina at Chapel Hill. Her interests include computational models of language-learning, in particular morphological learning, as well as models of linguistic categorization, and issues in theoretical morphology and morphophonology. Gergana Popova is a lecturer in linguistics at Goldsmiths, University of London. Her research interests are in theoretical linguistics, especially morphology and its interfaces with syntax and lexical semantics, as well as the analysis of text and discourse. Bert Remijsen is a linguist working at the University of Edinburgh. His main area of expertise is suprasegmental contrasts, e.g. tone, vowel length, stress, and voice quality. Earlier, Bert investigated such phenomena in the Austronesian languages Ma’ya and Matbat, and in the Caribbean Creole Papiamentu. Now he focuses on the Western Nilotic languages, in particular Dinka and Shilluk, where rich inventories of suprasegmental distinctions mark extensive morphological paradigms. Andrew Spencer is Professor of Linguistics at the University of Essex. His recent work focuses on lexical relatedness (including ‘mixed’ categories), and the stem as a morphological primitive, and the architecture of morphological models. Publications include Morphological Theory (Blackwells), Clitics (CUP, with A. Luís, 2012), and Lexical Relatedness (OUP 2013). Sabine Stoll is director of the Psycholinguistics Lab at the University of Zurich. Her work centres around cross-linguistic language acquisition research in first language learners with special emphasis on corpus analysis and method development. Her research programme focuses on learning mechanisms, which allow children to acquire maximally diverse languages. Recently Sabine co-edited The Acquisition of Ergativity (2013). Thomas Stolz is full professor of linguistics at the University of Bremen (Germany). His research interests include morphology, language typology, language contact, language change, and areal linguistics. Thomas has published extensively on topics related to languages such as Chamorro, Icelandic, Maltese, and Classical Nahuatl. He is editor-in-chief of the journal Sprachtypologie und Universalienforschung (STUF). Gregory Stump is Professor of Linguistics at the University of Kentucky. His areas of research interest include morphological, syntactic, and semantic theory and typology, with a particular focus on the form and content of inflectional morphology. Gregory is the author of Inflectional Morphology (CUP 2001) and, with Raphael Finkel, of Morphological Typology (CUP 2013); he is one of three co-editors of the journal Word Structure.

list of contributors

xxi

Jochen Trommer is Heisenberg fellow at the University of Leipzig and specializes in theoretical phonology and morphology, with a special focus on micro- and macro-variation in lesser-studied languages (Albanian, Uralic, Nilotic, Kiranti, and Algonquian). Currently his main interests are the learning of morphological segmentation and meaning, the role of tone in phonology and morphology, and the residue of non-concatenative morphology (polarity and subtraction). Matthew Walenski is Research Associate in the Roxelyn and Richard Pepper Department of Communication Sciences and Disorders at Northwestern University and an Adjunct Research Assistant Professor in the School of Speech, Language and Hearing Sciences at San Diego State University. Eva Zimmermann is a research assistant at Leipzig University. Her research is in theoretical phonology, theoretical morphology, and its interfaces. Currently, Eva works mainly on non-concatenative morphology (especially lengthening, subtraction, nonconcatenative allomorphy) and its analysis in phonological theory. An empirical focus in her work has been the non-concatenative morphology in several Native American Indian languages and the verbal agreement morphology in Kiranti and Broader Algic.

chapter 1 ........................................................................................................

INTRODUCTION ........................................................................................................

matthew baerman

1.1 Inflection

.............................................................................................................................................................................

Inflection is the expression of grammatical information through changes in word forms. For example, in an English sentence such as my cousin admires you, although the subject is clear from its placement before the verb, it is also indicated by the final -s on the verb, which tells us that the subject is third person singular, and hence cousin and not you, and also that the verb is in the present tense. This is fairly straightforward: we have an easily indefinable element (suffix -s) with an easily identifiable function (marking the presence of a third person singular present tense subject), but it seems to be in the nature of words that they may well follow their own rules, and that these rules may get quite a bit more involved. For example, the selection of forms in (1) from Chiquihuitl´an Mazatec, an Otomanguean language of Mexico, shows how difficult it may be to say where exactly the grammatical information is being expressed. Subject marking is distributed across the word form, involving the initial segments, for example hu- versus ˇcho-, the final vowel, e.g. -æ versus -e, and tone, for example high (marked here as superscript ‘1’) versus mid-high (superscript ‘2’). Further, each word in this little sample has a different pattern of alternation for all three of these elements. (1) Chiquihuitl´an Mazatec verbs, incompletive positive forms (Jamieson 1982: 166) hu1 bæ1 cˇ ho2 be2

‘I pull’ ‘you pull’

bu1 ya1 bo3 ye2

‘I return’ ‘you return’

hba3 nˇcæ1 hba2 nˇci2

‘I weed’ ‘you weed’

As impressive as such systems may be, it is perhaps an even more striking fact that, from the point of view of language as a whole, inflection is apparently superfluous. While every language operates by stringing together words, manipulating the shapes

2

matthew baerman

of words is further elaboration that many are happy to do without. This is probably what has lent inflection the curious and often contentious status it has within linguistics. On one view, inflection is just a particular instance of the more general lexical, semantic, syntactic, and phonological properties needed to characterize language as a whole, so that a word form such as admire+s is the syntactic combination of lexical units in the same way that the sequence my+cousin+admires is. Word-internal syntactic and phonological rules may differ from those that apply to other domains, but they are all cut from the same cloth. However, others see the behaviour of inflected forms— both the relationships that obtain between the forms of a single word, and those that obtain across different words—as potentially too divergent from the general set of linguistic modules for there to be a reliable mapping between them, and so postulate a distinct morphological component to grammar, subject to its own rules.1 Obviously, taking one or the other approach presupposes different models of the architecture of grammar, which in turn encourages different lines of enquiry. In the absence of an autonomous morphological component, questions of morphophonology and feature structure may assume prominence, whereas the assumption of an autonomous morphological component may foster a greater focus on questions of paradigm structure. But however they are interpreted, there is a fascinating and well-delimited set of facts out there to be explored. And if inflection is not shared by every language, it is itself a uniquely linguistic phenomenon, without obvious parallel in other communicative and symbolic systems. Inflection shares the larger realm of morphology with derivation, with which it has much in common, so it is important to distinguish the two in order to delimit the scope of the present volume. The basic difference of course is that inflection produces different forms of a word, while derivation produces new words. This is a rather slippery definition, as it presupposes that we can distinguish between different forms of a word and different words, but at some level we must all recognize that while the word cat is related to both cats and kitten, the nature of the relationship is different, both in terms of meaning and in terms of form. More precise attempts at a definition are of necessity multi-pronged. For example, Stump (1998: 15–18) offers the five diagnostics in Table 1.1. The first is in a way the most obvious: inflection preserves the lexical meaning while derivation changes it. The other diagnostics are essentially definitions of inflection, with derivation defined negatively: inflection involves a predetermined set of meanings (morphosyntactic values) which a word is required to accommodate, with the term derivation then applying to a looser constellation of morphological operations. The key point is that making a distinction between inflection and derivation leads to different expectations. With inflection, the set of word forms is predetermined, in that

1 While morphology obviously embraces more than just inflection, in practice only inflection is ever adduced as a reason for postulating an autonomous morphological component.

introduction

3

Table 1.1 Inflection versus derivation, per Stump (1998)

Lexical meaning and/or part of speech Obligatory Productive Semantically regular Closure (recursiveness)

Inflection

Derivation

Same Yes Yes Yes Yes

Different No No No No

a member of a given word class2 has certain duties to perform. For example, a verb X in English needs to be available for use in such sentences as She often X 1 or You X 2 yesterday. Where X = ski the relationship between X1 and X2 is transparent (ski-s, ski-ed), but the expectation that we should be able to put any verb into these contexts forces us to look beyond obvious formal resemblances, sanctioning such strange bedfellows as go and went. Derivation for its part has no preset list of job duties, and thus no binding obligations. The fact that we can derive duckling from duck does not sanction the expectation that there should be a comparable derivative of turkey (not for this author, at any rate), and a paraphrase or compound is good enough (say, turkey baby or turkey chick). As a consequence, the notion of derivation is in practice reserved for those instances where there is a formal resemblance between items that can be expressed as a morphological rule. Thus while the semantic relationship between puppy and dog is the same as that between duckling and duck, few would want to label this a derivation, and the status of gosling vis-`a-vis goose (in any synchronically meaningful sense) will depend in part on how much faith one puts in the morphological rules needed to relate them. Thus for any given language, inflection embraces a more-or-less closed system of interrelated functions and forms. In a sense, we can treat inflection as a discrete and self-contained slice of language as a whole, in the same way that we can treat language itself as a discrete slice of human cognition. This does not necessarily mean committing oneself to the notion that there is a component of language exclusively dedicated to inflection (or morphology more generally), any more than talking about language presupposes belief in an autonomous language faculty. And whatever status we accord an interface phenomenon such as morphosyntax—whether as syntax, morphology, or a domain of its own—it is evident that the existence of inflected word forms has contributed materially to the terms that we use to talk about it, and the concepts behind it. The following chapters are meant to address inflectional morphology from as many thematic angles as possible. Particular attention has been paid to taking examples from a wide variety of languages. This helps ensure that the topics covered are of broad 2 Word class in this sense is a morphological notion that may well cut across syntactic categories; for example, the verbal paradigm in Latin is traditionally construed as containing nominal forms, namely participles and gerunds.

4

matthew baerman

cross-linguistic relevance and makes for a lively and varied presentation. There are no chapters explicitly devoted to the exposition of particular morphological theories. This is intentional, and comes from the conviction that the most important difference between theories is the phenomena they choose to focus on, not the technicalities of their implementation. The approach taken here is therefore ecumenical, and the chapters come from various theoretical perspectives, as appropriate for the topic. We hope that this will be a useful and engaging resource for the study of what we find one of the most fascinating aspects of language.

1.2 About the Volume

.............................................................................................................................................................................

The first part covers the fundamental building blocks of inflectional systems. The morpheme (Chapter 2) has long served as the basic unit of morphological analysis, an almost unavoidable point of reference even within those schools of thought that reject it on principle. Anderson traces the history and differing conceptualizations of the notion, showing far greater divergence than is usually acknowledged. Features (Chapter 3) characterize the content side of inflection. Rather than treat each inflected form as expressing an isolated unit of meaning, we typically see them as manifesting particular values of higher level and possibly intersecting categories, such as case and number. The logic of feature systems is sometimes at variance with their formal expression, leading to some descriptive and theoretical challenges. Exponence (Chapter 4) describes how inflectional feature values are translated into morphological forms. A common assumption is that we can distinguish concatenative morphology, which involves the addition of discrete phonological units, from non-concatenative, which involves transformational operations or mapping onto preset templates. This chapter takes a more nuanced approach, using a finer-grained classification to integrate both types. The second part focuses on what is probably the most characteristic property of inflectional systems, paradigmatic structure: its nature, its variants, and its interfaces with phonology and syntax. The general concept of the paradigm is the topic of Chapter 5, which surveys the various ways it has been construed in different morphological traditions. Paradigms have assumed particular relevance in recent work which examines the implicational relationships that obtain between inflected forms, using insights offered by the application of information theory. The systematic variation between functionally equivalent but formally distinct alternative paradigms that occurs in many languages results in inflection classes. Chapter 6 addresses the theoretical, typological, and diachronic issues that they raise, considering the ways in which they may best be represented and accounted for. Inflection classes represent one dimension in which the mapping between function and form is not straightforward, since they seem to involve arbitrarily different patterns of inflectional exponence across

introduction

5

different lexemes. Cross-cutting this we also find deviations within the paradigm, surveyed in Chapter 7, manifested by such phenomena as syncretism, deponency, and defectiveness, each involving a less-than-straightforward relationship between morphosyntactic values and the forms that realize them. While this might be taken as evidence of purely morphological forces determining the structure of the paradigm, an alternative approach is to attribute some measure of these paradigmatic effects to phonology, as explored in Chapter 8. Chapter 9 steps beyond the strict confines of the inflected word form to look at periphrasis, where the inflectional paradigm appears to embrace multi-word constructions, including borderline cases involving clitics, whose status lies somewhere in between that of a free word and an affix. The third part covers change and variation over time. Chapter 10 surveys fundamental issues in diachrony, offering a typology of morphological and morphosyntactic change, as well as providing a picture of the larger context, including cycles of change and the role of other linguistic components as a triggering factor. Since inflectional morphology is generally resistant to borrowing, the topic of contact-induced change is relatively understudied. Chapter 11 presents the borrowing of inflectional morphology as the interplay between two factors: the scope of borrowing (inflected word forms or isolated inflectional material) and its influence on morphosyntactic categories already existing in the language. The fourth part covers computation, both as a tool for theoretical exploration and for practical applications. Chapter 12 looks at modelling inflectional structure computationally, as a way of clarifying the nature of inflectional operations and comparing alternative treatments. While these techniques involve explicit writing of inflectional rules in computationally tractable terms, machine learning (Chapter 13) can be used to explore more abstract questions about the nature of rules, the trade-off between rules and listing, and the role of linguistic universals versus learning biases in the development of inflectional structure. Although techniques of machine learning can be construed as modelling at least some aspects of human language acquisition, machine translation (Chapter 14) understandably takes a pragmatic approach to the issue, with inflectional morphology posing a particularly difficult challenge. In this quickly developing field we see an increasing convergence between rule-based and statistical methods, an interesting parallel to what we also see in purely theoretical treatments of inflection. Part V approaches the relationship of inflection to the human mind from two perspectives. Chapter 15 looks at the acquisition of inflectional morphology, with a particular focus on cross-linguistic variability. It argues for a step-by-step process of schema acquisition and extension, rather than a dual-route split between storage and rules. Chapter 16 explores the effects on inflection of a range of neurocognitive disorders. In contrast to the preceding chapter, it is suggested here that the evidence is consistent with a dual-route model of inflection, with damage to semantic memory affecting irregular (i.e. stored) morphology, and disorders affecting frontal/basal ganglia (related to motor functions) affecting regular (i.e. rule generated) morphology.

6

matthew baerman

Part VI is devoted to sketches of individual inflectional systems, illustrating a range of typological possibilities across a genetically diverse set of languages. Iha (Chapter 17) is a Trans New Guinea language spoken on the western tip of the island of New Guinea, and is particularly notable for the range of argument categories marked by agreement on the verb, with grammatically aligned suffixes (marking subject and object) interacting with semantically aligned prefixes (patient versus other). Pulaar (Chapter 18) is the westernmost dialect of Fula (Atlantic, Niger-Congo), spoken in Senegal, and is known for its heavy reliance on consonant mutation as an inflectional exponent. As with many other Niger-Congo languages, there is an extensive noun class system, and corresponding to this a large gender system. The conflicts in gender that would arise in conjoined noun phrases cause real difficulties, and are usually avoided. A striking feature of its morphosyntax is the relative tense, which involves a peculiar constellation of morphological and syntactic properties (e.g. inversion of verb–pronoun order), in a set of constructions which so far have defied precise characterization. Lithuanian (Chapter 19) is a Baltic language of the Indo-European family, known for a number of conservative features, including its complex nominal paradigms, involving both suffixal and prosodic alternations. A number of points of special interest are discussed here, including (i) the quasi-agglutinative local cases, calqued from the corresponding constructions in Finnic, (ii) the doubly problematic status of the reflexive, which is a borderline case between inflection and derivation on the one hand, and between clitic and affixal status on the other, (iii) evidentials, an unusual feature for an Indo-European language, and (iv) aspectual distinctions, which are only weakly grammaticalized and hence of uncertain status. Chamorro (Chapter 20), an Austronesian language of the Marianas islands (including Guam), stands out among the languages illustrated here in the categorial indeterminacy of its roots, which can be inflected as nouns or verbs, its ergative alignment, and the use of infixation and reduplication as productive means of inflectional exponence. Verbs in Murrinh-Patha (Chapter 21), a non-Pama Nyungan language of northern Australia, display a complex two-part structure, where lexical stems combine with members of a small set of highly irregular classifier stems, so that grammatical and lexical meaning are distributed across these two word forms. The resulting system is rife with discontinuous dependencies, multiple exponence, and the interspersal of inflectional and derivational material. Aymara (Chapter 22), eponymous member of the Aymaran family of Bolivia and Peru, represents a highly agglutinative, exclusively suffixing system. Nevertheless, it also displays a convincing instance of subtractive morphology, in that the accusative case is formed by apparent vowel deletion. Nen (Chapter 23) of the Morehead-Maro Family of Southern New Guinea, has a particularly elaborate system of distributed exponence in which verbal argument features are constructed through information spread across prefixes, suffixes, stems, and pronouns. Its aspectual system is based on the typologically unusual distinction of ingressive versus non-ingressive. Inflectional exponence in Shilluk (Chapter 24), a language of the West Nilotic branch of NiloSaharan (spoken in South Sudan), is characterized by stem-internal alternations of

introduction

7

length, tone, and ATR (advanced tongue route) values, along with stem-final consonant alternations. Since its affixal morphology shows a great deal of homophony or polyfunctionality, stem-internal marking plays an increasingly dominant role in the system.

Acknowledgements

.............................................................................................................................................................................

First and foremost I would like to express my appreciation to the contributors to this volume, who have done most of the work. The collaboration has been a rewarding one. I also thank Penny Everson for her invaluable assistance in the preparation of the manuscript. Finally, none of this would have been possible without the support of the European Research Council (grant ERC-2008-AdG-230268 MORPHOLOGY), which is gratefully acknowledged.

part i ........................................................................................................

BUILDING BLOCKS ........................................................................................................

chapter 2 ........................................................................................................

THE MORPHEME: ITS NATURE AND USE ........................................................................................................

stephen r. anderson

It is somewhat surprising that many surveys of the field of morphology—even quite comprehensive ones, like Spencer and Zwicky (1997)—devote scant attention to the nature of the morpheme, its ontological and epistemological status. They assume with remarkably little discussion that we know what morphemes are, and where and how to find them, and that the core problems of the field lie elsewhere. My intention here is to suggest that this optimism is misplaced. Jones’s (1962) monograph from which the present chapter draws its title makes an essential point about core theoretical notions in linguistics. An understanding of a term like ‘phoneme’ (in Jones’s book) must be grounded in more than just what linguists have written about its definition and place in the ontology of linguistic theory: we must also go to the literature and examine what they do with their phonemes, how the concept is actually used. This is at least as true for the ‘morpheme’, and it is necessary to explore both what people say morphemes are and how they in fact deploy them. I begin with some history, tracing the origins of the term and its development in theorizing through the structuralist period. Some basic problems with the classical conception are pointed out, problems which were quite familiar (but not substantively resolved) in the structuralist literature. Although this chapter appears in a handbook of inflection, I draw my illustrations both from inflectional and from derivational morphology. This seems reasonable, since general discussions of the morpheme have treated it as a concept equally applicable across this distinction. No issue would arise here, of course, for those linguists who deny the existence of a principled distinction between derivation and inflection, and no current proposal for ways to differentiate the two would appear to have the consequence that some of the formal situations to be discussed in Section 2.1.4 are predicted to arise in the one domain but not the other. I then consider the role of morphemes in early generative grammar, a context in which discussion of morphology per se was marginalized and its traditional domain subjected to a sort of ‘Partition of Poland’ between syntacticians and phonologists.

12

stephen r. anderson

The subsequent re-emergence of interest in morphology as a distinct field soon led to a divergence of theoretical views, much of which can be put down to differences of opinion about the nature and status of morphemes. Close examination of the practice of morphologists, however, suggests that the actual theoretical cleavages in this regard are not always as they might seem.

2.1 Some Background

.............................................................................................................................................................................

In linguistics, morphology is the study of words, and since the morpheme is at the core of morphology, we begin with the basic notion of a word. Although notoriously subject to differences of sense when seen from different perspectives, a word is fundamentally the unit of mutual association among several sorts of properties. Thus, the word cat in English has phonological properties (/kæt/), semantic properties (‘Felis catus, a small furry domesticated carnivorous mammal that is valued by humans for its companionship and for its ability to hunt vermin and household pests’), and syntactic properties (Noun, [+Common, +Count]). The connections among these are irreducible: that is, it is not possible, say, to associate the vowel /æ/ by itself with the ‘cute, furry’ part, or the final /t/ with its status as a common (not proper) Noun.

2.1.1 Origins: Signs, Basic and Otherwise As such, cat nicely fits de Saussure’s (1916 [1974]) notion of a minimal linguistic sign (even though Saussure did not include syntactic content in his account of individual signs). But Saussure realized that not all words are minimal signs. Even ignoring the problems presented by phonological words that include clitics together with their hosts (for which, see Anderson 2005), a word like unavoidable has a rather different character from that of cat. The phonology here is something like /ˌʌnəˈvoidəblˌ/; the semantics more or less ‘not possible to avoid’ and the whole thing an Adjective. In this case, parts of the phonology are connected to parts of the meaning, so the initial /ʌn/ represents the ‘not’ part, the final /əbˌl/ the ‘possible to’ part, and /əvoid/ the ‘avoid’ part. Such a word Saussure called a partially or relatively motivated sign. We might regard its parts as minimal signs in themselves, and of course these are the objects we teach our students to identify as morphemes in Linguistics 1. Considering the facility with which they can carry out that exercise after being given only a few examples, the notion seems to have a good deal of intuitive content. We could thus say that a first approximation to the notion of the morpheme is ‘a word or part of a word which constitutes an irreducible, minimal sign’. Saussure himself does not use the word morpheme (or its French equivalent, morphème) to refer to constituent parts of a relatively motivated sign, although the

the morpheme: its nature and use

13

term had been introduced as early as 1880 (as German Morphem) by Jan Baudouin de Courtenay, who defined it as ‘that part of a word which is endowed with psychological autonomy and is for the very same reason not further divisible. It consequently subsumes such concepts as the root (radix), all possible affixes, (suffixes, prefixes), endings which are exponents of syntactic relationships, and the like’ (Baudouin de Courtenay 1895 [1972]: 153; Stankiewicz’s translation). By the time of the lectures that form the basis of Saussure’s Cours de linguistique générale, he was quite familiar with Baudouin’s work, which he praised on various occasions. It is therefore noteworthy that he did not adopt—or at least mention— Baudouin’s designation for meaningful parts of a word. The potential significance of his failure to adopt this (or any other) distinctive terminology for such units is notable, in light of Wells’s (1947: 8) observation that ‘the term morphème was current in Saussure’s day, but with a specialized significance: the “formative” elements of a word (affixes, endings, etc.) as opposed to the root’. The usage Wells is referring to here was presumably that of Antoine Meillet, his students, and colleagues. In his French translation of Karl Brugmann’s 1904 Kurze vergleichende Grammatik der indogermanischen Sprachen, Meillet rendered Brugmann’s term Formans (Brugmann 1908), with the sense indicated by Wells, as morphème, and remarked in a letter to Baudouin de Courtenay that he had borrowed Baudouin’s ‘joli mot’, as he called it, for this purpose.1 In Meillet’s work, a language’s morphologie generally refers to its patterns of inflectional (and derivational) marking, and he uses morphème in various works to designate the formal reflections of morphological categories. For example, Meillet and Vendryès (1924: 148) say that ‘the principle of formation of Indo-European words is affixation: that is, to the element expressing the concept of an idea or an object (semanteme) are added the various elements (morphemes) marking the categories of words or their grammatical relations’ (my translation). Baudouin’s (1895) definition clearly intended a more general understanding of the notion of the morpheme, and this is the way it has been interpreted subsequently. Baudouin was a great coiner of neologisms, and his innovated word for the smallest indivisible component of a word was clearly based on the word ‘phoneme’, already in use for a minimal element in the analysis of sound.2 Saussure, on the other hand, maintained a view of the sign relation on which it holds between whole words3 and 1

See Mugdan (1986) for further discussion of this and other uses of ‘morpheme’ and its cognates in other languages by linguists from Baudouin to the present day. 2 It is worth noting that the word phoneme as used at that time had none of the connection with distinctiveness that it acquired in later structuralist usage. Phonème was originally suggested by Antoni Dufriche-Desgenettes as a term for a (basic) speech sound. Mugdan (2011) argues that Dufriche’s sense of the word involved a degree of abstractness beyond that of the physical speech sound, but not involving a criterion of distinctiveness. Saussure’s own use of the word phonème was rather idiosyncratic, but these matters need not concern us here. See Anderson (1985) for some discussion. 3 The ‘word’ is probably not the right unit to employ here, in any of its (diverse) standard senses, since much the same must be said about multi-word expressions whose meaning is not compositional. I ignore that complication here.

14

stephen r. anderson

their meanings, and so had less need for a way to refer to such a unit within words (contrary to what is usually suggested, as argued in Anderson 1985 and CarstairsMcCarthy 2005). Saussure would presumably have been familiar with Meillet’s usage, as well as Baudouin’s, and his failure to designate any sort of meaningful sub-parts of words with such a special term can plausibly be seen to be a consequence of his general failure to recognize such components as important linguistic objects in their own right. It is possible to see in the difference between Saussure’s and Baudouin de Courtenay’s usages the beginnings of a basic division in attitudes towards morphological structure that characterizes the field today. For Saussure, the ‘relative motivation’ of a sign like French poirier ‘pear tree’ resides in its relation to poire ‘pear’, a regular relation in form that is correlated with a regular relation in meaning. This is quite different from the description of poirier as composed of two pieces (morphemes), poire and -ier. Baked differs in form from bake, and the difference carries the significance ‘past tense’; but this is no different from the relation ran bears to run, which supports the same sense. The meaning of ‘past tense’ is linked to these (and other) ways of differentiating the related words, and not with a separable piece of one of the forms. Arguably, Saussure’s relational notion of morphological composition does not require a special term to designate the pieces of a partially motivated sign, and thus was the first version of what we can think of as a ‘rule-based’ conception of word structure as opposed to Baudouin de Courtenay’s ‘morpheme-based’ analysis.

2.1.2 The American Structuralist Morpheme A useful starting point4 for the discussion of notions of the morpheme in structuralist theory is the discussion in Bloomfield’s (1933) classic book. This is partly because Bloomfield’s notion of the morpheme served as the jumping-off point for later theorizing, but the book is also instructive for its occasional divergences between theory and practice. Presenting his views in the context of a thoroughly behaviourist set of rigorous procedures for discovering the elements of a linguistic analysis, Bloomfield often took his morphemes where he found them without requiring them to have emerged from the mandated procedures when those procedures failed to yield the desired answer. Bloomfield (1933: 161) defines a morpheme as ‘a linguistic form which bears no partial phonetic-semantic resemblance to any other form’, that is, a form that contains no sub-part that is both phonetically and semantically identical with a part of some other form. Unpacking this slightly in the commoner formulation as ‘a minimal same of form and meaning’, it is the requirement that phonetic and semantic resemblances be correlated, and it yields as ‘morphemes’ the elements that result when further division would destroy that correlation. 4

The account in this section largely follows that of Anderson (1992: ch. 3).

the morpheme: its nature and use

15

Taken literally, this definition leads to a variety of problems. One of these concerns the presence in a great many languages of ‘phonaesthemes’: sound-symbolic material such as the initial sl- of several English words referring to frictionless movement (slip, slide, slither, etc.); the initial gl- of words like glow, gleam, glitter, glimmer, glare, etc. referring to light emitted from a fixed source, or fl- in words like flash, flare, flicker referring to transitory light sources. Bloomfield (1933: pp. 244ff) enumerates a number of these (without really addressing the issue they pose for his definition of the morpheme), and they have been discussed often (if inconclusively) in the subsequent literature. What is at stake in these partial resemblances is a similarity in (reasonably) concrete semantics, and so it is unlikely that exactly parallel phenomena exist in the domain of purely grammatical inflectional morphology, but their bearing on the general notion of the morpheme remains. Bergen (2004) shows that these meaningful sub-parts of words are quite real for speakers, but linguists have not really known what to do with them. There is general agreement that they are not to be identified as morphemes, but the basis for excluding them is quite unclear. They have distinctive (if sometimes rather vague) semantics, correlated with distinctive phonological shape. It is not possible to write them off on the basis that the residue once they are subtracted is typically not a recurrent element itself—why is the decomposition of glimmer as gl+immer (cf. also shimmer = sh+immer, from a set also including shine) fundamentally more problematic than that of huckleberry as huckle+berry? Linguists are of one voice, however, that there must be a principled difference. Bloomfield calls these resemblances root-forming morphemes, thus treating them as a sort of morpheme, but others have generally wanted to find some analysis that does not have that consequence. Another difficulty is more technical, and served as the basis of subsequent elaboration. Bloomfield’s definition seems to assume that morphemes have a determinate phonological content, and as such is closer to the later usage of the term morph or allomorph. Later papers (e.g. Harris 1942; Bloch 1947; Hockett 1947; Nida 1948) refined the notion along the lines of the developing structuralist understanding of the phoneme. Just as phonemes came to be seen as abstract elements realized by members of a set of phonetic segments (their allophones), so morphemes were interpreted as abstract structural elements realized by members of a set of concrete phonological forms (allomorphs). Bloomfield’s actual practice is quite in line with this—he treats duke and duchess as sharing a morpheme with two alternants, even though it is hard to derive this analysis from his definition. The resulting view involves a commitment to several basic principles: (1) a. Morphemes are homogeneous, indivisible atomic units of linguistic form linking some component(s) of meaning with a set of mutually exclusive allomorphs that express it. b. Each morpheme has a determinate semantic content, and each allomorph has a determinate phonological form. c. Words are composed exhaustively of morphemes.

16

stephen r. anderson d. Each morpheme in a word is represented by one and only one allomorph; and each allomorph represents one and only one morpheme.

It was essentially in this form that American structuralists presented their generative successors with the notion of the morpheme, and with it, the field of morphology. Within morphology, this definition implies a clear division of labour, again modelled on practice in phonology. Allomorphy is the analysis of the paradigmatic conditions determining the range of possible variants of a single morpheme and the conditions under which each appears. The remainder of the study of morphology falls under morphotactics, parallel to phonotactics in being the analysis of the principles governing the syntagmatic distribution of morphemes in relation to one another.

2.1.3 European Structuralism My focus here is on theories of the morpheme in American structuralism, because the view that emerged there was the one inherited by later generative approaches to grammar. The linguists who were most influential in developing that picture came primarily from the anthropological tradition, with a focus on fieldwork procedures as the primary path to the study of language, and as a result, their idea of how to define terms like ‘morpheme’ was to specify a set of mechanically applicable procedures that would lead to the discovery of such elements. It is plausible to argue that they thereby fell victim to Bazell’s (1952) ‘correspondence fallacy’, one form of which is the assumption that a mechanical procedure identifying some objects that correspond to an intuitive, pre-systematic type will necessarily provide an adequate definition that completely reconstructs the traditional understanding of that type. In the first half of the twentieth century, when the American notion of the morpheme was being elaborated, the major linguists in Europe came from rather different backgrounds in philosophy or the philological study of various languages. For them, procedural approaches to basic theoretical constructs were less important than more traditional forms of definition that emphasized the direct cashing out of pre-systematic intuitions. In practice, the differences were somewhat limited: proceeding from their understanding of the morpheme as a minimal Saussurean sign, they arrived at much the same units in looking for the ‘morphemes’ of a language as their procedurally oriented colleagues across the Atlantic. Some differences in the extension of the term ‘morpheme’ did emerge. Martinet (1960), for example, distinguished morphèmes as units of grammatical meaning from lexèmes, units of lexical meaning, continuing the usage of Meillet described in Section 2.1.1. Morphèmes and lexèmes together constituted for Martinet the class of monèmes; but while there was a difference in the kind of meaning conveyed by the two types of monème, both were units of association between components of phonological form and components of meaning. Hjelmslev (1943) limits the use of morpheme to a unit of content (not form) corresponding only to inflectional categories, and not other meaningful elements (though

the morpheme: its nature and use

17

it is quite difficult to place Hjelmslev’s views on this, as on many other basic notions, in relation to those of other scholars). The restricted uses of the term by Martinet, Hjelmslev, and their colleagues, however, did not represent in themselves important differences between European and American understandings of the structure of complex words. The Swedish linguist Adolf Noreen (1854–1925) also used the word morpheme (or rather, its Swedish counterpart morfem) in a somewhat idiosyncratic way. For Noreen, a morpheme was an expression with a unitary associated meaning, defined recursively so that affixes, words, and phrases all count as morphemes: dog, -s, dogs, big dogs, etc. are all morphemes on this view, which goes together with a usage of phoneme that allows it to apply to arbitrarily long segments of phonetic form, rather than to an individual segment. He was followed in this usage of ‘morpheme’ by a few linguists in Scandinavia (including Valentin Kiparsky, father of Paul Kiparsky), but the notion did not catch on elsewhere. A difference in practice between Europe and America concerned the role of meaning. As American structuralism solidified around an essentially behaviourist conception of language, meaning tended to be marginalized or disregarded altogether. Morphemes were supposed to have not only a form but an associated meaning, but a common assumption was that meanings were intrinsically unavailable for study in themselves, and so American linguists tended to be content with the observation that the difference between one morpheme and another corresponded to some difference in meaning without feeling a need to say much about what the actual meanings involved might be. European linguists of the period were much more interested in descriptive semantics, and this had consequences for their work in morphology. Roman Jakobson, in particular, took the principle that a morpheme has a meaning to the conclusion that this meaning ought to be unitary, even in cases where some apparent diversity appeared to be present. Jakobson’s (1936) study of the category of Case accordingly attempted to propose a unitary common meaning (‘Gesamtbedeutung’) for each of the formally distinct nominal cases of Russian, with results that have been debated ever since. This line of research produced a variety of studies in response, but since the work has formed a part of the research tradition more in semantics than in morphology, it will not be pursued further here. The related industry within morphological analysis of decomposing morphosyntactic features into component values also does not bear directly on our concerns here.

2.1.4 Termites in the Foundations of the Structuralist Morpheme The picture structuralist morphology led to is one on which morphemes (construed as sets of allomorphs) have on the one hand a meaning and on the other hand

18

stephen r. anderson

an instantiation in a particular phonological alternant. The principles in (1), then describe the (ideal) relation of meaningful morphemes to the allomorphs composing the phonological representation of the word form. This idealization is rather at odds with the empirical facts, however, and in this section I survey some of the ways in which it is inaccurate. Most of these were recognized already by Hockett (1947); somewhat unaccountably (to a later reader), Hockett seemed to feel that giving each of the problematic cases a name sufficed to resolve the difficulties they appear to pose for the theory of the morpheme. The phenomena of circumfixation (e.g. Indonesian kebebisan ‘freedom’ cf. bebis ‘free’; kedatangan ‘arrival’, cf. datang ‘come’: Sneddon 1996: 35ff.) and infixation (e.g. Sundanese ɲaraian ‘to wet’ (pl), sg ɲaian: Robins 1959; or Koasati hocífn ‘smell-2sg’, cf. hófn ‘smell-3sg’: Kimball 1991) demonstrate that the phonological expression of a morpheme is not necessarily continuous and indivisible, since (part of) the expression of some other morpheme may interrupt it. Some authors (e.g. Corbin 1987) have attempted to argue that genuine circumfixes do not in fact exist, and that apparent instances can always be decomposed into the combination of an existing prefix and suffix. These arguments are not convincing, however, for languages of the Indonesian type (where the meaning and distribution of a circumfix may be unrelated to properties of any prefix and/or suffix); and in any event, the formal problem is the same for infixation (where the stem into which the infix is inserted becomes in effect a circumfix). Some generalization of the phonological form of allomorphs, perhaps along the lines suggested by McCarthy (1981) is thereby indicated, and such an extension does not appear to compromise the essential content of the traditional notion of the morpheme in fundamental ways. Closely related is the problem of multiple exponence: systems in which the same category is marked in multiple places within the form. An example is provided by negation in Muskogean languages such as Choctaw (Broadwell 2006: 148ff.). In this language the affirmative form iyalittook ‘I went’ corresponds to the negative akíiyokiittook ‘I didn’t go’. There are several separate aspects of this latter form that mark negation: (a) one set of subject markers is replaced by another (-li is replaced by a-); (b) k- is prefixed to the stem; (c) -o(k) is suffixed; (d) the stem vowel carries an accentual feature of length; and (e) the optional suffix -kii has been added. Such multiplication of markers is similar to the circumfix case, but some of the morphemes involved may be linked to other things as well (and may have independent motivation). Verbal agreement in many languages involves multiple marking of the same content. A simple example (for which see e.g. Aronson 1982) is the Georgian verb movdivar ‘I come’, where both underlined v’s indicate first person subject (cf. modixar ‘you come’). Another example is provided by class marker (gender) agreement in Batsbi5 (Harris 2009), a Northeast Caucasian language as illustrated in (2).

5

Also known in the literature as Tsova Tush.

the morpheme: its nature and use

19

(2) tišin c’a daħ d-ox-d-o-d-an-iš old house pvb cm-destroy-cm-tr.prs-cm-evidI-2pl.erg ‘You (pl) are evidently tearing down the old house.’

The verb here contains three separate instances of the marker d, all triggered by agreement with the noun class or gender of the object c’a ‘house’. Particularly exuberant expression of the same verbal argument in multiple places in the form is characteristic of the Kiranti languages of Nepal, such as Limbu (van Driem 1997). The histories of these languages appear to involve multiple repetitions of the coalescence of an inflected auxiliary with a preceding (independently inflected) main verb, leading to synchronic states of affairs in which two, three, or even more markers refer to the same argument: for example Limbu dza ŋpətə ‘I’m going to eat’, where both ŋ and the final schwa are markers of first person subject agreement. The Papuan language Skou (Donohue 2003) provides another instance in which multiple markers correspond to a single verbal argument, and where the history that has led to this situation can be reasonably established. Such examples demonstrate not only that multiple exponence is possible in agreement systems, but also that historical change is not constrained by a requirement that it not result in systems with this property. The problems posed by various forms of discontinuous expression of the same content material for the principles in (1) would seem to be largely mechanical, but some authors (e.g. Halle and Marantz 1993; Steele 1995) have asserted, on the basis of a commitment to strictly discrete and local expression of morphosyntactic properties along the lines of (1), that multiple exponence can never occur in the morphological system of a natural language. As discussed in Anderson (2001), such a theoretical position must be rejected or modified in light of the many clear cases of this phenomenon in the languages of the world. There are other, more fundamental difficulties. Consider first the existence of empty morphs, morphological material that does not correspond to any part of the meaning of the form in which it appears. For example, in several Algonquian languages, inflectional person marking prefixes such as Cree ki- ‘1st person’, ni- ‘2nd person’, and o‘3rd person’ are followed by an empty element -t- when added to a vowel-initial noun or verb (Wolfart 1973; cf. ospw¯akan ‘pipe’, ot¯ospw¯akan ‘his pipe’; ast¯aw ‘(he) puts (it)’, nitast¯an ‘I put it’). This element is the product of historical rule inversion: in earlier stages of Algonquian, stem-initial /t/ was deleted except when ‘protected’ by a personal prefix. This led to t/∅ alternations which were later resolved by treating all of the relevant stems as vowel initial, and introducing a /t/ precisely when such a stem (including both original t-stems and those originally beginning with a vowel) is preceded by a personal prefix. Derivational examples include the underlined material in English crime/criminal, page/paginate, sense/sensuous, habit/habitual/habituate; Spanish madre/madrecita (cf. co-madre/co-madrita); or Russian slog ‘syllable’, slogovoj ‘syllabic’ (cf. odno-složnyj

20

stephen r. anderson

‘monosyllabic’). Similar formations are found in vast numbers of languages, and compromise the notion that morphemes always have semantic content (or else that words are always exhaustively analysable into morphemes). A distinct but related problem is posed by superfluous morphs. These are elements that can be shown to have content in some of their occurrences, but where that content is inappropriate in others. The most widely cited example of this situation is the feminine formative that appears in Romance adverbs like French doucement ‘sweetly, gently’. Feminine agreement is quite unmotivated in French adverbs, and appears only as the historical reflex of their origin in a construction with an adjective modifying the feminine noun m¯ens (ablative mente). This is by no means an isolated example: consider, for example, the th and vowel change in English lengthen, strengthen, or the t in heighten. Note that for example lengthen means ‘to make long’ and so we should expect ∗ longen, ∗ strongen, ∗ highen (like shorten, weaken, shorten). The formation of strengthen, lengthen, heighten on the basis of strength, length, height rather than the semantically more appropriate strong, long, high is motivated not by its apparent morphological composition but rather by a phonological condition discussed by Siegel (1974), to the effect that causatives and inchoatives in -en can only be formed from bases ending in an obstruent. In the resulting form, we must say either that the additional ‘morpheme’ has no meaning, or else that this meaning is somehow disconnected from the meaning of the whole word. The complementary problem is presented by zero morphs, instances where some aspect of a form’s content is not reflected at all in its form. The poster child of such formation is the Russian genitive plural of nouns like dáma ‘lady’, GenPl dam. Treating such cases by positing a morpheme with no phonological content, but only a meaning, is a time-honoured form of analysis, but that does not mean it is really consistent with the classical understanding of the morpheme. Jakobson’s (1939) eloquent defence of such zeros as the way to analyse cases like that of the Russian genitive plural certainly demonstrates that forms with no overt marker can be opposed to other forms within the same paradigm just as affixed or otherwise marked forms can. That does not, however, show that the appropriate way to do that is to posit a ‘minimal same of form and meaning’ with no overt form. Describing this situation by appeal to a ‘zero morph’ does not provide a solution to the problem, but only a name for it. It seems simply to be a fact that many words have morphological properties that are not reflected in any way in their surface shape. The presumed separability of morphemes is compromised by the existence of overlapping morphs. In Breton, for example (cf. Press 1986), e dad ‘his father’ contrasts with e zad ‘her father’ (cf. tad ‘father’). Here the possessor is marked by the preceding e and the particular mutation of the initial consonant that is associated with it. As a result, the initial segment of the noun is simultaneously part of the exponent of the possessor and that of the stem. The limiting case of this is what Hockett (1947) called portmanteau morphs, like French au = à le where the two elements are coextensive and coincide completely in a single undecomposable form.

the morpheme: its nature and use

21

Similar are cumulative morphs, such as the suffix -¯o of Latin am¯o, ‘I love’ which realizes in a single segment the categories of person, number, tense, mood, and voice that are parcelled out among multiple components of other members of the paradigm such as am¯abam ‘I loved, was loving’. Another clear example of such cumulation of categories in a single marker is the Finnish Nominative Plural ending -t, as in talo/talot ‘house/houses’ (compare Partitive taloa/taloja, where case and number are expressed separately). The classical conception of the morpheme suggests that every word can be exhaustively decomposed, on the one hand into a sequence of meaning elements, and on the other into a sequence of phonological substrings, such that the relation between the constituent elements of the two sequences is one-to-one (and ‘onto’ in both directions, for the mathematically inclined). The facts just surveyed, among others, suggest that this does not correspond to the general case, and thus that words cannot be required to be analysed as sequences of morphemes in this sense. Comparable sorts of problem with the ideal agglutinative picture are just what characterize ‘inflectional’ languages in the classical typology. When we incorporate the necessary emendations as codicils to the theory of the morpheme, the general picture becomes quite unconstrained as far as the relation between phonological and semantic form and thus loses much of its original appeal. There is also another, quite different, set of problems for the classical picture of words as made up of morphemes. Because words are presumed to be partitioned into discrete, separable morphemes, the basic form of this theory holds that all of the phonological content relevant to signalling morphological content should be uniquely assignable to concrete segments. But sometimes the aspect of a form’s phonological shape that indicates some aspect of its content does not consist of segments (or parts of segments) at all. Trivially, this is true for ‘zero morphemes’, since they have no phonological content, but more fundamental difficulties have long been recognized. Quite widespread in many languages are various formations that can be subsumed under the general heading of apophony, including ‘Umlaut’, ‘Ablaut’, ‘gradation’, and others. Consider the relations among the forms of English strong verbs, such as sing, sang, sung (cf. also song), or (American English) dive, dove. Their analysis poses a classical problem that was discussed in great detail in a landmark paper by Hockett (1954). When we ask what the correct analysis of for example sang is, several possibilities present themselves, none of them entirely satisfactory from the point of view of the traditional morpheme. We might say that the past tense morpheme here has a zero allomorph, and that sang is a predictable allomorph of sing that appears before this past tense zero. This sort of analysis was deplored early on by Nida (1948): it involves saying that the thing we cannot see, the zero, is what signals that the verb is past tense, while the thing we can see, the vowel change, is analysed as a mechanical concomitant of this. The result does not correspond to any plausible intuition about how form and content are related. Alternatively, we might treat the vowel /æ/ as the past tense marker, but then we are forced to say that verbs like sing/sang have a Semitic-like consonantal stem (/s—ŋ/) and

22

stephen r. anderson

unlike others, a distinctive present tense marker /i/. Apart from its basic implausibility, this account has trouble with verbs like dive that show (for many North American speakers) two alternative past forms, weak (dived) and strong (dove). The analysis in question has the consequence here that the two past tense variants are related to identical present tense forms that nonetheless represent totally different structures. Another possibility is to say that the past tense marker in sang is a ‘replacive’ morph (/i/→/æ/), not a piece of phonological content in itself but an operation on the content of the stem to change its vowel. It is quite difficult to see how the procedures of segmentation and classification appealed to in structuralist theories could ever discover an element of this sort. More importantly, perhaps, it is difficult to see how such an operation of replacement is consistent with the notion that a morpheme is an association of meaning and form, with some concrete piece of the phonological form signalling the corresponding meaning of past tense. Here what we really want to say is not that some aspect of the shape of sing indicates past in itself, but rather that past is indicated by the relation between sing and sang. Examples abound in language where it is such a relationship between forms that indicates their respective morphological content, and not some discrete affix added to one or the other. Apart from tense in strong verbs and a few nouns with residual Umlaut plurals (mouse/mice, (wo)man/(wo)men, etc.), English is not often thought to offer many instances. In fact, though, they are easy to find if one includes relations other than inflectional ones: consider pairs such as, sell/sale, blood/bleed, food/feed, etc., where vowel differences serve morphological functions, and believe/belief, prove/proof, speak/speech, bath/bathe, breath/breathe, glass/glaze (provide with glass), use ([jus], noun)/use ([juz], verb), in which consonant changes operate in the same way. Consonantal alternations marking inflectional categories are found in some languages. For example, Uralic languages often show a system of consonant gradation that depends on syllable structure: thus, in Finnish at the beginning of a short, closed syllable geminate stops become single, and single stops become (the reflexes of original) voiced segments. In the Saami languages, final nasals have been lost, and as such there is no longer an overt suffix to mark the genitive (typically homophonous with the accusative). The gradation alternations originally associated with the addition of a final nasal in these cases remain, however, as the only marker in some paradigms. Thus, North Saami (Sammallahti 1998) has alternations such as guolli ‘fish’, gen/acc guoli; giehta ‘hand’, gen/acc gieđa, etc. Fula (West Atlantic) has a system in which every stem potentially occurs in three distinct shapes that differ in the category of their initial consonant. These form three grades, a ‘continuant’ grade, a ‘stop’ grade, and a ‘nasal’ grade. Without going into the complex phonological details of this system (see Arnott 1970 and Anderson 1976 for discussion), the choice of one grade or another constitutes part of the agreement system both for nouns and adjectives and for verbs. Each of the more than two dozen noun classes of the language is associated with a specific grade, and ‘the stem-form appropriate to each class is as much a grammatical feature of the class as the concordmarking suffixes’ (Arnott 1970: 93). Similarly, the choice of a nasal-grade form of

the morpheme: its nature and use

23

the verb stem marks a plural or post-posed subject as opposed to a preposed singular subject. Particularly difficult for the notion that morphological markers are always to be identified with some distinct substring within a word are cases where the marking instead is by subtraction. A class of nominals in Icelandic (e.g. hamr ‘hammering’ from hamra ‘to hammer’) have been widely cited in this regard. In these forms, we can show phonologically from the distribution of vowel length and other properties that the noun is formed from the infinitive by deleting the final /–a/ that marks infinitives (see Orešnik and Pétursson 1977 for discussion). In the structuralist literature, the most widely cited example of subtractive morphology was the supposed formation of masculine adjectives in French by deleting a final consonant from the feminine. This is almost certainly not the right analysis of this case, however; the feminine is instead formed from the masculine by adding a final schwa, preserving a final consonant that would otherwise be lost, as suggested by the orthography (and supported by other considerations: see Anderson 1982a and much other literature on the phonology and morphology of French). Although this example is probably not valid, other instances of subtractive morphology surely do exist. In such cases, there is no requirement that the deleted material be a ‘morpheme’ in its own right. In the Icelandic nominals referred to in the previous paragraph, the deleted -a is in fact the infinitive suffix, but since all infinitives (the bases for the formation) end in this element, it is not possible to distinguish the phonological and morphological characterizations of the deleted material. In other cases, however, the deleted material clearly does not have any distinctive morphological value. In the Muskogean language Alabama, for instance, plural forms of verbs are made by deleting the rhyme (nucleus plus coda) of the final syllable of the stem: balaa-ka ‘lie down (sg)’; bal-ka ‘lie down (pl)’; batat-li ‘(I) hit once’; bat-li ‘(I) hit repeatedly’; kolof-li ‘(I) cut once’; kol-li ‘(I) cut repeatedly’ (Broadwell 1993). In Huichol (Uto-Aztecan; cf. Elson and Pickett 1965: 48f.), perfective verbs are formed from imperfectives by subtracting the final syllable, which is not itself a distinct marker of any category: nepizeiya ‘I saw him (and may see him again)’, nepizei ‘id. (for the last time)’; pïtiuneika ‘he danced (and may start again)’, pïtiunei ‘id. (and will not again)’. In all of these cases, a category is marked precisely by the absence of some phonological material we would otherwise expect, and not by the presence of some marker. Another way to mark morphological categories formally (but without adding material to the form) is by metathesis. The best known proposed example of this is from the Salish language Klallam, where a sort of imperfective form of the verb is made by inverting the order of the stressed vowel and a preceding consonant: ˇckw ú–t ‘shoot’, ˇcúkw –t ‘shooting’. Comparable formations appear in a number of related languages of the family. In some instances, there is controversy about the correct analysis, but there is good reason to believe that in at least some of the languages involved, morphologically conditioned metathesis is definitely present (see Anderson 2004 for discussion and references). At least in Saanich, the incompletive category in question clearly

24

stephen r. anderson

involves a morphologically conditioned rule, and the relation between completive and incompletive is marked precisely by the relation between CCV and CVC. Another apparent example from outside Salish is suggested (Mel’ˇcuk 1997: 297) in the Kartvelian language Svan, where transitive causative verbs are related to intransitives by means of a similar CCV↔CVC difference: cf. li-deg ‘go out’, li-dge ‘put out’; li-k’wes’ ‘break (intr)’, li-k’ws’e ‘break (tr)’. Such signalling of morphological content by means of the re-arrangement of existing material, rather than by the addition of a distinct marker, is difficult to reconcile with the traditional notion of the morpheme. Similar difficulties are presented by marking based on exchange relations as in Diegueño (Yuman; Langdon 1970), ł y ap ‘burn (sg)’, ł y a:p ‘burn (pl)’; sa:w ‘eat (sg)’, saw ‘eat (pl)’. Here the singular/plural relation is signalled by an interchange in the value of vowel length. Also problematic for the classical morpheme are relations based on chain shifts for example Saami (‘Lappish’), where as already noted, the genitive is related to the nominative base in many nouns as geminate stop to simple voiceless stop, or simple voiceless stop to voiced (/pp/→/p/, /p/→/b/). As in the other types discussed here, there is no constant content to the formal expression of the category (here, genitive). If we wish to maintain that words are exhaustively composed of morphemes, and that morphemes in their turn are discrete units representing the association of a part of the word’s sound with a part of its sense, all of these types of morphological marking pose problems. From their examination, it becomes clear that not all components of a complex word’s content are indicated by distinct affixes in the way this picture would suggest. In fact, problems of this type are not limited to the domain of phonological expression. Consider: if every morpheme is an association between some form and some meaning, it ought to be the case that adding a morpheme involves adding some form, and as we have seen, that is not always the case. But it ought also to be the case that adding a morpheme entails adding some meaning, and that is not uniformly the case either. Again, there are zero cases: the empty morphs, where an added piece of form corresponds to no added meaning. There are also the superfluous morphs, where the morpheme has a meaning, but this meaning does not contribute to the sense of the form in which it appears. More interestingly, there are also semantic analogues of subtractive morphs, where adding a formal ‘morpheme’ actually removes some of the semantic content of the base form. Consider some pairs of Icelandic verbs distinguished by the presence of the ending -st, such as gleðjast ‘rejoice’, gleðja ‘gladden (tr.)’; kveljast ‘suffer’, kvelja ‘torture (tr.)’; týnast ‘be, get lost’, týna ‘lose’, and many others (see Anderson 1990; Ottósson 1992 for discussion). Here the -st marker has the effect of detransitivizing the basic verb, removing from its meaning the components characterizing a causative relation between an agent and some state of affairs. We could represent the semantics of the transitive bases here as something like (cause x, (become (P y))) (e.g. ‘sbj causes obj to become happy, to suffer, to be lost, etc.’). The addition of the ending -st has the effect of removing the highest predicate

the morpheme: its nature and use

25

(cause x,y) from this structure (and also deleting the corresponding argument position and/or θ-role from the syntax). Phonological and morphological arguments show that -st is added to the base; syntactic and semantic ones show that the form with -st has less semantic content than the related transitive (i.e. what is involved is not simply something like the binding of the agent variable with an impersonal operator, or the like). The conclusion is that the addition of the -st ‘morpheme’ here has a subtractive effect on the meaning of the resulting form, parallel to the phonologically subtractive effect seen in other examples earlier in this subsection. Similar examples can be found in many languages, where intransitive verbs are morphologically more complex than corresponding causative transitives. This fact is typically concealed in analyses by glossing the additional marker as something like ‘DeTrans’ so that it appears to add something to the meaning, but when we ask what that something might be, it turns out that the effect of the marker is actually to eliminate some of the semantic content of the base. Semantically subtractive markers are just as problematic for the notion of words as uniformly composed of traditional morphemes as phonologically subtractive ones. As in the case of phonaesthemes, the fact that such examples rely on the lexical semantics of the items in question suggests that corresponding cases will not be found in the domain of inflection, but they nevertheless bear on the general tenability of the classical conception of the morpheme. If we take seriously the evidentiary value of the morphological types surveyed in this section, it is evident that the most we can say in general about the analysis of words is approximately as in (3). (3) A linguistic sign relates a word’s content and its form. The content can be divided into its syntactic properties and its meaning; each of these can be further analysed. The form can be analysed into phonological segments (organized into syllables, feet, etc.). The relation between content and form may be partially systematic. This is more or less equivalent to Saussure’s original recognition of the existence of partially or relatively motivated signs, and offers no particular privileged status for a unit like the morpheme as traditionally construed. Nonetheless, linguists continue with disconcerting regularity to regard analyses such as the decomposition of unavoidable into un+avoid+able as if it provided a perfectly general model of word structure.

2.1.5 The Fate of the Morpheme in Post-Structuralist Grammar Many of the difficulties noted in Section 2.1.4 for the concept of morpheme were quite familiar to structuralists, and formed the basis for a somewhat contentious literature devoted to their resolution. Nonetheless, the basic idea that words were to be regarded as exhaustively composed of morphemes, where these were elementary units linking phonological form with meaning, was not really challenged. This was felt to be secure

26

stephen r. anderson

on the basis of the wide range of cases for which it provided a perfectly satisfactory basis for analysis. The difficult cases would have to be resolved somehow, perhaps by tinkering in various ways with definitions, but were not taken to pose fundamental difficulties for a notion that generally worked well in daily life. Of course, the view that the earth is flat is also consistent with a wide range of observations, provides a satisfactory basis for most of our projects, and generally works well in daily life. The rise of generative theories of language in the 1950s and 1960s involved a rejection of many of the tenets of structuralism, but nonetheless built on structuralist notions in many ways. As a student of Zellig Harris, Chomsky inherited some of the conceptual apparatus of American structuralism, and as a student of Roman Jakobson, Halle brought the perspective of that particular version of European structuralism. The innovative character of the new approach to language resided in its insights in the areas of syntax and phonology, and these were the domains that occupied scholars’ interests. Morphology in itself was not something that generative theory was much concerned with. Besides, the concrete domain of morphology seemed vanishingly small. Recall that the nature of the classical morpheme suggests a division of morphological description between two subfields: allomorphy and morphotactics. But once generative phonology had abandoned the notion of a phoneme based on surface contrast, and with it the distinction between phonemics and morphophonemics, it appeared that the vast majority of variation in the shape of morphological units would be subsumed under phonology, leaving nothing in this area for morphologists to do but list the unpredictable (suppletive) forms. And from at least as early as Chomsky’s (1957) analysis of ‘Affix Hopping’ in English it was assumed that the syntax could manipulate internal constituents of words, leaving little in the way of morphotactics as residue. This last move bears a mild irony, since in structuralist times it was assumed that syntax itself was nothing but the ‘morphotactics’ of larger and larger domains: generativists and structuralists were thus agreed on the unity of syntax and morphotactics, though they differed on where the action was in describing this set of phenomena. In any event, from the generative point of view there did not seem to be much independent substance to the study of morphology. With only minimal interest as a focus of independent attention, the morpheme’s appeal as a basic descriptive unit was not subjected to close examination, and so the structuralist conception was imported more or less unmodified into generative theories.

2.2 The Morpheme in Current Theories

.............................................................................................................................................................................

Beginning with work in the 1970s, generative linguists gradually came to see that the abandonment of morphology as a distinguishable aspect of the structure of language

the morpheme: its nature and use

27

was probably ill-advised. Regularities of variation in shape that are sensitive to morphological content appeared to follow principles that fell outside those of phonology, and the internal organization of words seemed to have quite different grounds from the organization of words into phrases and sentences. As the reductionist waters of the initial results of generative inquiry receded, a lost continent of morphology came back into view. The most natural assumption for most linguists was that this was a land populated primarily by morphemes of the traditional sort, and that interpretation has persisted at least in the rhetoric of adherents of several distinct theoretical positions.

2.2.1 The Rediscovery of Morphology The retention of the structuralist ontology is clear in the papers that provided the early charter for the investigation of morphology within generative grammar. Halle (1973: 3f.) notes that ‘the assumption has been made quite generally that a grammar must include a list of morphemes as well as rules of word formation or morphology’ and that ‘the list must include not only verbal, nominal, and adjectival roots but also affixes of various sorts’. Halle’s paper does not challenge the assumption that the items on this list are essentially associations between phonological and semantic content, though he does add, in accord with the much greater importance attributed to syntax at the time than in structuralist work, that ‘[i]t is all but self-evident that in the list of morphemes the different items . . . must be provided also with some grammatical information. For example, the entry for the English morpheme write must contain the information that it is a verbal root, that it is a member of the “non-Latinate” portion of the list (it is by virtue of this fact that it is allowed by the rules of word formation to combine with certain affixes and not with others), that it is among the small class of verb stems that undergo the so-called “strong” conjugation, etc.’ Halle’s account of morphology in his 1973 paper is primarily a treatment of the ways in which the morphemes on this list can be combined into larger structures to make words. He does depart from previous assumptions in assuming that some of the information associated with words is linked to them as wholes, and not contributed directly by any of their constituent morphemes. This necessitates a list of the words of the language, its ‘dictionary’, which is separate from the list of morphemes and serves as a filter on the output of word formation, sanctioning some combinations and disallowing others while adding word-specific information, such as the idiosyncratic sense of recital as referring to a concert by a soloist (and not simply an act of reciting), whether a particular word is or is not subject to certain phonological processes such as trisyllabic shortening, whether the existence of a given word does or does not block other comparable formations, etc. Halle’s (1973) architecture for morphology seems to involve a non-trivial amount of duplication between the effects of rules of word formation and the content of the ‘dictionary’, and was not widely pursued. His attention to words (and not

28

stephen r. anderson

simply morphemes) as the locus of significant properties was, however, taken up and expanded in another of the foundational works on morphology of the period, Aronoff (1976). The influential theory presented in that book was based on the notion that word formation rules, rather than simply organizing elements from a list of morphemes into larger structures, operated on entire words as basic elements, relating them to other words. Aronoff ’s word-based morphology comes closer to being based on what I have described in Section 2.1.1 as Saussure’s view of the sign, but does not in the process dispense with a reliance on morphemes. He does point out yet another class of problems for the traditional morpheme: prefix–stem combinations in English like prefer/confer/presume/consume etc., where there is apparently internal morphological structure but where neither the prefix nor the stem can be assigned a semantic interpretation on its own. Nonetheless, the words in a language’s lexicon that serve as the inputs and outputs of word formation rules are assumed to be structured concatenations of rather traditional morphemes, once provision is made for such cases. These morphemes are presumed to be identifiable by the rules of the morphology, and to be accessible in the structural descriptions and structural changes of those rules. For example, Aronoff invokes a set of rules of truncation, one of which serves to suppress the -ate of navigate in the formation of navigable. He proposes that such rules have a quite specific form: they always have the function of deleting a specified morpheme (here, /-At/) in the environment of members of a list of other morphemes (here including /-əblˌ/). The validity of this analysis6 is not at stake here: what matters to the present account is the reliance it places on the conception of words as composed of rather traditional morphemes, even though the word formation rules themselves operate on words as wholes. Other work in the ensuing development of a generative approach to morphology was quite explicitly based on classical morphemes as the basic structural unit. Selkirk (1982) for example treats the analysis of word structure as the extension of syntactic principles (such as ‘X-theory’) to word internal domains, where they would serve to organize morphemes in hierarchical structures entirely comparable to the syntactic organization of phrases. Another important paper of the time, Williams (1981), is explicit in suggesting that morphological relatedness is to be reconstructed as the sharing of morphemes, where these are organized into structures that derive their overall properties through a notion of ‘head’ deriving from syntactic theory. Williams (1989) argues that these structures are grounded in the same notions of ‘head’ and ‘projection’ 6 In fact, following the insightful observations of Corbin (1987) concerning truncation in French, Anderson (1992) proposes that what is truncated in such a case is precisely not a morpheme in the classical sense. On that account, truncation is a component of word formation processes (such as the one that forms adjectives in -able from verbs) that serves to accommodate material in words borrowed from another language that has morphological significance in that language, but where the morphology involved is opaque in the borrowing language. This issue is not directly relevant here, but it should be noted that the existence of truncation phenomena, on this account, does not furnish evidence for the significance of morphemes.

the morpheme: its nature and use

29

that operate in syntax, although the actual substantive parallels he cites seem rather inconsequential. Lieber (1980, 1992) pursues a very similar line, construing morphology as the ‘syntax of words’ in a manner that recalls, with allowances for differences in the overall theoretical context, the structuralist account of syntax as essentially the ‘morphotactics of words’. All of this work is fundamentally based, explicitly or implicitly, on the assumption that words are to be seen as structured organizations of morphemes, where these basic elements are discrete units associating phonological, semantic, and syntactic information in essential accord with the structuralist principles in (1). Problems with this notion of the sort discussed in Section 2.1.4 are not generally discussed. Where examples are considered in which morphological information is indicated by some aspect of a form other than a segmentable stem or affix, this is attributed to mechanical manipulation of the word’s phonological shape by members of a vaguely specified class of ‘readjustment rules’ or ‘morphologically conditioned phonological rules’, whose operation is triggered by the presence of significant zero elements, perhaps in conjunction with ‘grammatical’ features of individual words or morphemes. Nida’s (1948) objection noted in Section 2.1.4 to the procrustean nature of such analyses is not generally addressed, or acknowledged.

2.2.2 Distributed Morphology and the Morpheme Halle (1990) introduces a shift in the notion of the morpheme that serves as a bridge to later theories, in particular that of Distributed Morphology. In this paper, Halle assumes that morphemes are of two sorts, distinguished by their phonological nature. One type, concrete morphemes, are characterized by ‘a single fixed underlying phonological representation’. These are to be distinguished from abstract morphemes, which ‘do not have a fixed phonological shape’ and thus ‘lack a phonological U[nderlying] R[epresentation] in the vocabulary entries’. Both types of morpheme are represented in the terminal strings of syntactic representations, but differ in that while concrete morphemes have a phonological shape, abstract morphemes are simply bundles of features ([Plural], etc.). They are then provided with a phonological interpretation through a set of spell-out rules: for instance, the English morpheme [Plural] is spelled out as /∅/ following one set of nouns (sheep, man, moose, etc.); as /i/ after another set ending in /us/ (which is deleted in the plural), etc., and as /z/ by default elsewhere. While the examples of abstract morphemes that he considers are inflectional in nature, Halle resists the suggestion that the distinction between concrete and abstract morphemes is equivalent to that between stem and derivation, on the one hand, and inflection on the other. This is because some inflectional elements (e.g. Spanish -mos ‘1pl verbal ending’) have a constant shape. On the other hand, some variation in shape does occur in concrete morphemes (e.g. man/men), but in that case it is described by the operation of ‘readjustment rules’.

30

stephen r. anderson

Halle’s distinction between concrete and abstract morphemes is problematic on various grounds, but what is significant about it is the proposal that at least some morphemes constitute the basic components of words from the point of view of the syntax, but are only supplied with phonological form at a late point in the derivation through the operation of the ‘spell-out’ process. This conception is extended to all morphemes (stems and affixes) in Halle and Marantz’s (1993) presentation of the theory of Distributed Morphology. That theory implements the view that only the morphosyntactic properties of an element, and not its specific semantics or phonology, are relevant and visible to the syntax. ‘The terminal elements of the tree consist of complexes of grammatical features. These terminal elements are supplied with phonological features only after Vocabulary insertion. . . . Although nothing hinges on this terminology in what follows, we have chosen to call the terminal elements “morphemes” both before and after Vocabulary insertion, that is, both before and after they are supplied with phonological features’ (Halle and Marantz 1993: 114). Within this theory, the traditional analysis of morphemes is partitioned among several parts of the grammar. Allomorphy, or the variation in shape of morphemes, is separated into three components. Phonologically conditioned variation, such as the variation in the shape of the English regular plural among the forms [-s], [-z], and [-əz] is governed by the normal phonology of the language. Unpredictable, suppletive variation (e.g. the fact that the plural of ox is formed with the ending [-ən]) is described by context sensitive rules of Vocabulary insertion. Finally, some variation such as stem vowel Ablaut in English strong verbs (e.g. sing/sang/sung) is described by morphologically conditioned (‘Readjustment’) rules manipulating the phonological form introduced by Vocabulary insertion. Morphotactics, in contrast, is said to be described by the syntax. Morphemes (in the abstract sense of Halle) are distributed, both within and across words, by syntactic rule, and then phonologically realized by Vocabulary insertion. This picture would appear to predict that phonological units (introduced through Vocabulary insertion) would bear a straightforward, one-to-one relation to grammatical units arranged by the independently motivated rules of the syntax. If that were the case, we would expect the relation of form to content to be essentially the same as that envisioned by structuralist views based on morphemes conforming to the assumptions in (1). This is not, however, the case, as a result of the presence of additional structure in the model. Halle and Marantz assume a level of representation (‘M[orphological] S[tructure]’) which intervenes between the output of the syntax and the process of Vocabulary insertion. A variety of operations can result in differences between the syntactically motivated structure and MS: these include insertion, deletion, or movement of morphemes; combination (of two types, ‘Fusion’ and ‘Merger’) of two or more morphemes into a single unit, or ‘Fission’ of one morpheme into two; and copying of features from one morpheme node to another. Despite the claim (Halle and Marantz 1993: 121) that the manipulation of syntactic structure to produce MS occurs ‘only in highly constrained and fairly well understood ways’, the result is that the two can in principle be arbitrarily different from

the morpheme: its nature and use

31

one another. As a result, if we take the output of the syntax to correspond to the representation of content, and MS to the (schematic) representation of form, the relation imposed between the two is no narrower than the formulation in (3): partially systematic, but not discretely localized as required by (1). This does not in itself, of course, argue that the theory is incorrect or misguided, but it does show that it does not result in analytic units that correspond to the traditional notion of a morpheme. Although ‘morpheme’ is used in two quite different senses (one for the abstract, phonologically uninterpreted objects that serve as terminal nodes in the syntax, and the other for the result of phonological interpretation of the elements of MS), neither articulates the same understanding as any of the traditional uses of the word.

2.2.3 Morphology without Morphemes If the rather loose connection between form and content expressed by (3) is an accurate characterization of the structure of language in general, the presumption of a tight link between the two domains that is implicit in the classical notion of the morpheme is misguided, and morphological analysis should proceed on some other basis. A useful classification of theories due to Stump (2001b) separates contemporary morphological theories on the basis of a difference that can be seen as hingeing on their attitudes towards the traditional morpheme. Lexical theories, on Stump’s analysis, are those where associations between (morphosyntactic) content and (phonological) form are listed in a lexicon. Each such association is discrete and local with respect to the rest of the lexicon, and constitutes a morpheme of the classical sort. In contrast, inferential theories treat ‘the associations between a word’s morphosyntactic properties and its morphology’ as ‘expressed by rules or formulas’ (Stump 2001b: 1). This allows the systematicities foreseen in (3) to be expressed, while not requiring the exhaustive one-to-one matching of form and content presumed by (1). This distinction functions in combination with another to provide a substantive typology of theories. Orthogonal to the difference between lexical and inferential theories is that between incremental theories, on which a word bears a given content property exclusively as a concomitant of a specific formal realization; and realizational theories, on which the presence of a given element of content licenses a specific realization, but does not depend on it. In these terms, the view which is most congenial to the traditional morpheme as the locus of form and content in close association is a lexical incremental one, like that of Lieber (1992). Distributed Morphology is characterized by Stump as lexical, in that it assumes a listed set of form–content associations, but realizational, in that the relation between overall content and overall form is looser than that presumed by a strictly morphemic analysis. Inferential-realizational theories include those of Matthews (1972), Zwicky (1985), Anderson (1992), and Stump (2001b) himself.

32

stephen r. anderson

Inferential-realization theories represent, in effect, the complete abandonment of the traditional concept of the morpheme. In such a picture, a grammar includes a lexicon of basic forms linking semantic and morphosyntactic content with phonological form: these correspond to the ‘semèmes’ of Meillet and Vendryès (1924) or Martinet’s (1960) ‘lexèmes’. The formal exponents of grammatical content, however, the ‘morphèmes’ of these writers, are not elements taken from a lexical list, but rather the consequence of the application of modifications in form induced as the effect of a system of rules or relational formulas.

2.3 Conclusion

.............................................................................................................................................................................

Several fundamental distinctions in linguistic analysis are thus seen to be bound up in the usage by linguists of the basic term ‘morpheme’. One of these corresponds to the basic nature of the sign relation as this was introduced into the study of language by de Saussure (1916 [1974]). Although virtually undiscussed in the literature over the past century, the question of whether this relation holds between the form and content of whole words on the one hand, or more locally of minimal internal constituents of words, is a basic one. If the interpretation offered in this chapter is correct, this differentiated the views of Saussure from those of Baudouin de Courtenay and many who came after them, and persists today in the difference between lexical and inferentialrealizational theories. Across time, an overt indication of positions in this regard has been the attitude taken towards the role of the morpheme as represented by the role such a term plays in a theory’s ontology. A second, related difference can be traced back to that between Baudouin de Courtenay’s (1895 [1972]) use of ‘morpheme’ to include roots as well as affixes, vs. that of Meillet or Martinet for whom the word referred only to the markers of grammatical information, and not to basic lexical elements. Today that distinction corresponds to the difference between lexical theories of morphology, which assume that the lexicon contains all such elements of either sort, and inferential theories, which treat grammatical information as marked by grammatical mechanisms distinct from the insertion of lexical material. The word ‘morpheme’ is one of the most basic terms in linguistics, one which students are expected to control almost from the beginning of their study of the field. Linguists of many persuasions use the word freely, if only as a descriptive convenience, even when their theoretical commitments are not consistent with the idealized picture of word structure inherited from our structuralist forebears. We assume that both the intension and the extension of the term are virtually self-evident, but it turns out on closer examination to hold the keys to some of the deep questions we can ask about the nature of language. One of these, indeed, is whether or not there is any such thing as a ‘morpheme’.

the morpheme: its nature and use

33

Acknowledgements

.............................................................................................................................................................................

I am grateful to Joachim Mugdan and Paul Kiparsky for help with the history of the term morpheme. Matthew Baerman, Andrew Carstairs-McCarthy, and Joachim Mugdan have provided useful comments on a draft of the present chapter.

chapter 3 ........................................................................................................

FEATURES IN INFLECTION ........................................................................................................

greville g. corbett

3.1 Why Features?

.............................................................................................................................................................................

If we are analysing a small text or a substantial corpus, or even trying to describe a whole language, it makes sense to separate off and describe manageable parts of the problem. We find instances like English hats, where the comparison with hat leads us to analyse hats as consisting of the lexical meaning (piece of headgear) and the more abstract grammatical meaning plural. We note further that hat is also analysable as having the same lexical meaning, but being singular. Hat and hats contrast in NUMBER ; they show the singular and plural values of the feature NUMBER . Features like NUMBER are the normal means of modelling linguistic information of this type. They are particularly valuable where cross-cutting generalizations are found (as here where grammatical meaning cross-cuts lexical meaning). We can now generalize across other lexical items: women is like cats in being plural, though this is marked in a different way. We can go further and generalize across parts of speech: English determiners also distinguish NUMBER (this ∼ those) as do verbs (wears ∼ wear), though there is an interesting difference here, to which we return in Section 3.2. We can see generalizations across languages: many have a NUMBER feature which bears comparison with that of English. There can be differences as to which nouns differentiate NUMBER, there may be other feature values involved (such as dual), but there are also commonalities. Further discussion of the nature of features can be found in Corbett (2012).

3.2 Features for Different Components

.............................................................................................................................................................................

Features are of various types. There are the straightforward ones which apply within a single component. For instance, phonological features such as ±NASAL or ±APPROXIMANT are restricted to the phonological component (see Clements 2009

36

greville g. corbett

for a clear overview of phonological features). Thus it is claimed that syntax is phonology free (Zwicky and Pullum 1983; Pullum and Zwicky 1988); we do not find syntactic rules making reference to ±NASAL; there cannot be a rule of the type: ‘verbs beginning in a +NASAL segment are clause initial’. In addition there are features which apply across different components. Svenonius (2003: 376) nicely terms these ‘interface features’. The first to be discussed were morphosyntactic features (Matthews 1972: 162), which have a role in syntax and in morphology. These will figure large in our discussion; another key type of interface features is the morphophonological features, which naturally straddle morphology and phonology.

3.3 Describing Morphology with Features

.............................................................................................................................................................................

In order to lay out the features which may be required to describe morphological systems, it will be helpful to start with an abstract and idealized schema, and only then concentrate on real examples. The ‘language’ described will have some lexical items with accidental resemblances to English; the semantics can be assumed to be similar, but the morphological properties will prove rather simpler than those of English. Consider these related forms in our hypothetical language: play ∼ played, show ∼ showed; there are many more like this. The forms in –ed are used for reference to the past, while those without the suffix are for the present. As arguably in English, we have a regularity which affects only the meaning and the morphology, hence a morphosemantic feature of TENSE.1 Of course, there will be adverbs that are felicitous with one or other tense value, but that is a matter of semantic compatibility. This situation is what we might have expected, in that the information concerning tense is found ‘where it belongs’, namely on the item to which it most closely relates. Even in this simplest case, there is more that we may need to say. As in real English, hypothetical English does not allow *boyfriended ‘former boyfriend’, *nexted to the window ‘formerly next to the window but since moved’ nor *pinked ‘formerly pink, but subsequently dyed another colour’. That is, there is a condition on the morphological part of the regularity: it applies only to items of the right part of speech, namely verbs. There are apparently similar features which additionally have a role in syntax. Suppose our language has agreement in number. If there is agreement of the verb with the subject in number, this means that we have information about the number of the subject ‘in the wrong place’, namely on an item to which it does not directly relate 1

For real English one might wonder about sequence of tense rules as being syntactic; first these are not straightforwardly acquired as are normal rules of syntax but need instruction, and second they can be treated as a semantic problem (Higginbotham 2008).

features in inflection

37

(the point alluded to in Section 3.1). This number marking is determined by syntax, specifically by a rule of agreement. In this case NUMBER is called a morphosyntactic feature: it is an interface feature since it is involved in a rule of syntax and a rule of morphology. However, it still has a semantic side (morphosyntactic features need not match semantics completely, but equally they always reflect semantics to a considerable degree).2 The morphosyntactic features should be distinguished from the morphosemantic ones for the conceptual reason just given, and since they are different typologically: morphosyntactic features are more limited in number, and they tend to correspond more closely across languages. However, many linguists do not distinguish them and use ‘morphosyntactic feature’ as a general cover term to include also the morphosemantic features. In our hypothetical language, the forms for tense and those for agreement might be constant across all verbs. Imagine the plural marker is -z, which gives forms such as these: run (sg) versus runz (pl). But now we find other instances of the following type: rop (sg) robz (pl), sak (sg) sagz (pl) and many more. In fact looking further we find that this language has no final voiced stops. We treat this as a general rule of phonology. The phonological rule applies through the morphology as elsewhere, referring to the phonological feature ±VOICE, and there is nothing the morphological component need say about it. Our hypothetical language is indeed simple, but there is nevertheless more work to be done to characterize its morphology. We need to establish a list of the features and of their values. We then need to work out the rules of assignment to these feature values: what counts as past? Is there a today’s past distinct from a more remote past? Does ‘one and a half ’ require a singular or plural? Does the experiencer of a verb of perception stand in the dative, the affective, or the nominative? Next we need to establish the behaviour of forms which have complex featural specifications. For instance, if we have a verb which is morphosemantically past and morphosyntactically plural can both specifications be expressed: is the form played-z (play-pst-pl) or play-z-ed (play-pl-pst)? Or are the two specifications expressed cumulatively by some quite different means, for instance play-n (playpst.pl)? We also need to determine whether there are any interactions between features and their values: it might be, for instance, that GENDER is expressed in the singular but not in the plural, or indeed that agreement in NUMBER is found in the present but not in the past. Let us review the ground we have covered. To describe an inflectional system we need an account of the morphosemantic and morphosyntactic features which determine the different forms of lexemes. We shall need phonological features to describe the resulting forms, but in the simplest situation there will be no addition to the phonological

2 Hence these features would be more accurately but cumbersomely named ‘morpho-syntacticosemantic features’.

38

greville g. corbett

rules required elsewhere in the language. It is now time to tackle a real system. Consider the paradigms of two Russian nouns. (These are presented in a transliteration of the standard orthography, which is largely morphophonemic.) (1) Paradigm of two Russian nouns žurnal ‘magazine’ singular plural nominative žurnal žurnaly accusative žurnal žurnaly genitive žurnala žurnalov dative žurnalu žurnalam instrumental žurnalom žurnalami locative žurnale žurnalax

komnata ‘room’ singular komnata komnatu komnaty komnate komnatoj komnate

plural komnaty komnaty komnat komnatam komnatami komnatax

The CASE and NUMBER values assumed here can be justified relatively easily, though there are further values, and we return to one of these in Section 3.5.1. Nevertheless, we should ask how it is that we come to such an analysis. This was a major concern of the members of the Set-theoretical School (for which see van Helden 1993). Specifically, since žurnal has the same forms, in the singular and in the plural, for the case values nominative and accusative, why do we suggest that there are two case values here? The method is set out more fully in Zaliznjak (1967/2002: 36–42), but see also Goddard (1982), Blake (1994: 29–30), or Corbett (2012: 75–9). We take different syntactic contexts (such as ja vižu . . . ‘I see . . . ’), and collate the forms which fit appropriately into these contexts. If we had only the evidence of the first noun žurnal ‘magazine’, we would have to say that we had no evidence for different values of nominative and accusative. However, when we put komnata ‘room’ into contexts with slots for subjects and for objects, these provide evidence for distinct feature values (the traditional nominative and accusative). Hence we could claim that the two different instances of žurnal have different case values. Of course, there is more to it than that. The context must also be semantically constrained, and we may require alternative contexts in order to allow natural readings for different semantic classes of noun. In the straightforward instances we find that the expected features and values are established. Often, however, less clear instances emerge too. An underlying assumption here is the principle that syntax is ‘morphology free’ (Zwicky 1992: 354–6). There are two sorts of justification for this assumption. First, we aim for simple rules of syntax (such as: ‘X governs the accusative’), and these rules do not select for different classes of noun. That is, we do not expect to find syntactic rules referring to morphological information (for instance, ‘verb X governs the accusative of nouns of like komnata but the nominative of nouns like žurnal’). And second, in the instances which at first appear challenging for the principle, we find

features in inflection

39

that analyses respecting it prove more insightful (Corbett and Baerman 2006; Corbett 2009a). In order to maintain this view, that of morphology-free syntax, we are willing to have žurnal as the realization of the feature specification nominative singular and of the specification accusative singular. In other words the form žurnal is syncretic, realizing both the nominative singular and the accusative singular specifications. Morphosyntactic features are the key to inflectional morphology. Following our earlier discussion we see that they cross-cut lexical meaning. They also provide a classification which is orthogonal (cross-cutting) both to part of speech classification and to each other. As we saw, NUMBER may be relevant both to nouns and verbs, as in English. And in Russian, NUMBER and CASE are similarly orthogonal to each other. If we return to (1) above, the layout of the table with NUMBER as one dimension and CASE as another is more than just tradition. We need to be able to make statements about each feature, independently of the other. For example, the number value of the relative pronoun is determined by the number value of its antecedent (though the antecedent’s case value is not relevant). The preposition k ‘towards’ governs the dative case value (and has no ‘interest’ in the number value of the governee). And yet, the situation is often not as clear-cut as this picture might imply. Morphosyntactic features have ‘penumbras’, where the situation is less neat; they have values whose status is less clear. Often this is evidence of past changes. Similarly there are relations between forms which appear to be phonological, and yet do not achieve the regularity expected of phonological rules. We return to these issues in Section 3.5. First, however, we should tackle the main outstanding problem: I have suggested that the different forms in (1) are not different in morphosyntactic terms (which respects the principle of morphology-free syntax). How then are they to be accounted for?

3.4 Morphological Features

.............................................................................................................................................................................

The main issue is clear from (1) above. If we have established the morphosyntactic features and their values for a given language, this may be insufficient information for us to predict the forms of every lexeme. Specifically we cannot say for Russian what the marker of the accusative singular is. We can give the marker of the accusative singular only in respect of a given lexeme or set of lexemes. On the other hand, we do not need to give each form for each morphosyntactic specification separately: there are inter-dependencies between them. To make this clearer, it is worth moving on to a fuller account of the noun morphology of Russian (2):

40

greville g. corbett

(2) The major inflectional classes of nouns in Russian3

sg

nominative accusative genitive dative instrumental locative

pl

nominative accusative genitive dative instrumental locative

gloss members (nearest 50)

i žurnal žurnal žurnal-a žurnal-u žurnal-om žurnal-e

ii komnat-a komnat-u komnat-y komnat-e komnat-oj komnat-e

kost′ kost′ kost-i kost-i kost′ -ju kost-i

iv4 bolot-o bolot-o bolot-a bolot-u bolot-om bolot-e

žurnal-y žurnal-y žurnal-ov žurnal-am žurnal-ami žurnal-ax ‘magazine’ 20,700

komnat-y komnat-y komnat komnat-am komnat-ami komnat-ax ‘room’ 13,600

kost-i kost-i kost-ej kostj-am kostj-ami kostj-ax ‘bone’ 3,950

bolot-a bolot-a bolot bolot-am bolot-ami bolot-ax ‘marsh’ 5,750

iii

It is clear from this fuller picture that we cannot say what, for example, the form of the accusative singular is, without knowing the noun in question. But equally, the inflections are not distributed in a ‘pick-and-mix’ fashion across the lexical items: if we know one form, thishasimplicationsforotherforms. Infact, thenounsofRussiancanbeanalysedas falling into the four groups in (2), which we term inflectional classes (see also Stump, this volume). Traditional terminology has ‘declensional classes’ for nouns and ‘conjugation classes’ for verbs; we will treat them together as inflectional classes. Thus we look at an inflectional class from two perspectives: its forms and its members. In terms of forms, we ask what they are, how far they differ from the forms of other inflectional classes, and how far they can be predicted; for instance, in (2) above, the instrumental singular komnatoj is a good predictor of other forms, while the instrumental plural komnatami gives no further information (see further Blevins, this volume, on interpredictability of forms). In terms of members, we ask which items belong in which class, and again whether this is predictable. For instance, if in Russian a noun denotes a female human there is a strong probability that it will inflect like komnata (apart from effects due to animacy). We can get a picture of the distribution of the nouns over these four main classes by calculating the membership of around 44,000 nouns; (2) gives the number in each class, to the nearest fifty, calculated from Zaliznjak (1977). 3 As noted earlier, we give the forms in transliteration of the standard orthography, which is largely morphophonemic. Palatalization of the preceding consonant is indicated alternatively by ′ or by j. 4 Our four examples represent the vast majority of Russian nouns, but there are some further variants and exceptions, for which, see Timberlake (1993: 837–41) and there are several hundred indeclinable nouns. Most nouns, as in our examples have a single stem, but a minority have stem modifications.

features in inflection

41

The intersection of these two properties—a shared set of forms and a specified list of lexemes—produces the type of situation where linguists use features. We can have a morphological feature inflectional class, with various values. For now we can say that žurnal ‘magazine’ belongs to inflectional class I, and komnata ‘room’ to inflectional class II. There has been a certain amount of empty rhetoric about the need for such features. The point is simply that we need a way of modelling the fact that lexemes can inflect differently for the same morphosyntactic feature values; this can be done through a morphological feature, or whatever else is used as an equivalent. Consider for instance, Network Morphology accounts, which take full advantage of default inheritance (Corbett and Fraser 1993; Brown and Hippisley 2012). In such an account, a lexeme has various patterns of sharing with other lexemes, including those in different inflectional classes. For žurnal ‘magazine’, the fact that the dative, instrumental, and locative plurals are in turn žurnalam, žurnalami, and žurnalax is not specific to that inflectional class, it is information shared by practically all inflected nouns of Russian. In fact, the amount of information that needs to be specified for inflectional class I is just that the nominative singular is the bare stem, and that the genitive plural consists of the stem plus the inflection -ov. All the remaining forms have a wider distribution and can be inferred from elsewhere in the network of the inflections of nouns. In such an analysis, the inflectional class feature value functions as a hook to link the individual lexical entry into an inheritance network. The data in (2) illustrate the obvious fact that different items may realize the same morphosyntactic specification with different forms. But there can also be different patterns of realization: as we noted earlier, inflectional class II has a unique form for the accusative singular, while in the other inflectional classes there is no such unique form. Equally, the inflectional classes are not entirely different. Some morphosyntactic specifications have the same realizations (see, for instance, the dative plural). Some patterns are shared too: inflectional classes II and III both have syncretism of dative and locative singular. The picture presented in (2) covers many thousands of nouns, and so treating them with inflectional class features or equivalent makes good sense. However, as we noted earlier, what is unique to a particular class is the pattern (not all the forms are unique to the class). The fact that some forms are shared across inflectional classes means that on the one hand the child has fewer forms to learn, while on the other the task is more difficult, since this very sharing means that certain forms give the learner little indication as to the rest of the item’s paradigm. The picture presented in (2) is representative, in broad outline, of many similar systems. But the picture can be radically different. Thus in Burmeso, a language of the Mamberamo River area of Western New Guinea, for which, see Donohue (2001), discussed in Corbett (2009b), the verbal inflectional classes have forms which are fully distinct. The effect is that any one form allows a complete prediction of the full paradigm.

42

greville g. corbett

We have concentrated on the most important type of morphological feature, namely inflectional class features. These are used to account for differential realization of features by segmental material, as in our example, and also by prosodic means. Other interesting but less important proposed types of morphological features are discussed in Corbett and Baerman (2006).

3.5 Patterns of Past Meanings and Past Forms ............................................................................................................................................................................. Changes in both grammatical meaning and form leave behind patterns which may be synchronically more (or less) justified from outside the inflectional system.

3.5.1 Patterns of Past Meanings If a morphosyntactic feature goes into decline, that can leave forms which are difficult to describe. The Slavonic family had a number system, consisting of singular, dual, and plural. In most of the modern languages the dual has been lost, but it has left a remarkable set of remnants. Thus in Russian, it has left forms normally but not always identical to the genitive singular. The remnant forms are severely limited in two ways. First they are governed just by the numerals dva ‘two’, tri ‘three’, ˇcetyre ‘four’ (also oba ‘both’, poltora ‘one and a half ’, pol ‘half ’ in compounds), provided they are in the nominative case (or the accusative and are quantifying an inanimate). Second, there is a unique form just for a handful of nouns. For most nouns, the form matches the genitive singular as given in (2). For example, dva žurnala ‘two magazines’, where žurnal is in the genitive singular. Here the form is just as in cena žurnala ‘the price of the magazine’. The small number of nouns which have a special form in these circumstances are all in inflectional class I; they are: ˇcas ‘hour’, šag ‘step, pace’, šar ‘ball, sphere’, rjad ‘row’, sled ‘footprint’, and the unique form is not used obligatorily for all of them. The unique form is distinguished by stress only. Thus dva ˇcasá ‘two hours, two o’clock’ shows the special adnumerative, in contrast with okolo ˇcása ‘about an hour’ with the normal genitive singular. Thus we need an extra feature value to cover these various survivors of the old dual.

3.5.2 Patterns of Past Forms The remnants of past forms are seen in morphophonological alternations (which can also be modelled with features). A clear set of examples comes from Russian, as presented in (3), based on Timberlake (1993: 835).5 5 A somewhat more detailed account is given in Timberlake (2004: 82–4), where he terms them morpholexical alternations. See that source for the alternations in (3) which are not discussed in the text. For an implementation within Network Morphology, see Brown (1998).

features in inflection

43

(3) Morphophonological consonantal alternations in Russian basic grade p, b, f, v, m t, d, s, z k, g, x n, r l

alternation 1 plj , blj , flj , vlj , mlj cˇ , ž, š, ž cˇ , ž, š nj , rj lj

alternation 2 pj , bj , fj , vj , mj tj , dj , sj , zj cˇ , ž, š n j , rj lj

alternation 3 no change no change cˇ , ž, š no change lj

Take the following instances: plakat′ ‘cry’ (infinitive) pisat′ ‘write’ (infinitive)

plaˇcu ‘I cry’ (1 singular present) pišu ‘I write’ (1 singular present)

The infinitive forms have no alternation here: the consonant involved is one of those from the first column, those in the basic grade. For verbs of the type we are considering, the forms of the present tense have a consonant alternation: they take the forms from the second column (alternation 1). Historically these present tense forms trace back several centuries to changes known as the first palatalization of velars and the iotation of dentals and labials (Schenker 1993: 68–70). We should consider the phonological part of this alternation. First the consonants involved form phonologically motivated groups. This becomes clearer as one scans the different outcomes in (3) above: the groups which share behaviours are the labials, dental stops, velars, and resonants. And second, there is an obvious phonological patterning to the alternations: thus where š is the alternating form of s, the voiced pair works in the same way (ž is the alternating form of z). However, this is not just a matter of phonology. At the simplest level, there is no phonological rule of Russian requiring, for instance, that s be realized as š before u. We can see this in a verb of a different type: nesti ‘carry’ (infinitive)

nesu ‘I carry’ (1 singular present)

Thus we need to specify where an alternation will occur: it is not a matter of automatic phonology. Importantly too, it is not sufficient to specify that an alternation will occur, we need to specify which alternation will be found. As (3) shows, one and the same consonant can alternate in different ways in different morphological environments. For instance, s alternates to š in the examples we have seen. Elsewhere it can mutate to a palatalized variant (sj ). The morphological part of these morphophonological alternations, then, consists of two specifications. First there is the environment in which the alternation occurs: for some verbs it goes through the entire present tense, for others it involves the first singular only, for others the second and third persons singular and the first and second persons plural. And second, we need to specify which alternation is required; this corresponds to the columns in (3) above. The theoretical point

44

greville g. corbett

should now be clear: we have a generalization (a mutation type) which cross-cuts the consonants involved and the morphological environments in which the particular alternation is found. This is exactly the situation which can be modelled by a feature, with the values mutation 1, mutation 2, mutation 3. It is possible, given enough ingenuity, to write some of these changes into the phonological representation, using various diacritics. But we should not deceive ourselves: however we represent these alternations, they are the remnants of phonological rules which are no longer productive. And attempts to treat the Slavonic alternations as phonology miss the implicational relations between the patterns, as pointed out by Spencer (2012a: 237–9). The alternations survive in the morphology, in environments which have to be specified, and which may now signal morphological oppositions in some instances.

3.6 The Ultimate Morphological Default

.............................................................................................................................................................................

As we saw earlier, the shape of the paradigm is determined in large part by the morphosyntactic features. In (2) we had the dimensions of NUMBER and CASE, and these are what the syntax requires. We start from the assumption that the morphological paradigm will match this morphosyntactic requirement, and it often does, to a large extent. The fact that there is more than one set of forms answering the requirement points to the need for inflectional classes. Many of the sources of particular interest in inflectional morphology arise precisely from mismatches between the morphosyntactic requirement and the morphological realization. Thus syncretism, as discussed in relation to the Russian accusative, is precisely the situation where there are two morphosyntactic descriptions corresponding to single morphological form. This natural assumption of matching between morphosyntax and morphological paradigms is allowed for in Network Morphology: there is a lexemic hierarchy, dealing with an item’s morphosyntactic patterning and a morphological hierarchy, which is concerned with the realization of forms, and these two are the same by default.6 The interest is in the nature of the overrides. In general, the overrides are not sufficiently serious to make one question the basic matching between the two featural descriptions. However, sometimes the mismatch is dramatic, and this leaves the key issue of what the morphological feature set looks like in such instances. The Baltic language Latvian will help us address the issues. Latvian has the following undisputed case and number values. 6 For the comparable notions of content paradigm and form paradigm in Paradigm Function Morphology, see Stump (2012).

features in inflection

45

(4) Latvian noun paradigm: galds ‘table’ (from Veksler and Jurik 1978: 25)

nominative genitive accusative dative locative

singular gald-s gald-a gald-u gald-am gald-¯a

plural gald-i gald-u gald-us gald-iem gald-os

However, this paradigm is insufficient for the needs of the syntax. Veksler and Jurik (1978: 25), from whom the forms in (3) are taken, also include an instrumental, which is governed by the preposition ar ‘with’. There is no unique instrumental form, rather it is syncretic with the accusative in the singular and the dative in the plural. This is not something special about this class of noun; the same pattern of syncretism runs right through the language, including the personal pronouns. If we do not recognize the instrumental, then we have a preposition, ar ‘with’, which takes different case values according to whether the governed element is in the singular or the plural. There is a similar problem with the following examples (Veksler and Jurik 1978: 87; see Corbett 2010 for fuller references): (5) Gr¯uti dz¯ıvot bez hard live.inf without ‘It’s hard to live without a friend.’

draug-a friend-sg.gen

(6) Gr¯uti dz¯ıvot bez hard live.inf without ‘It’s hard to live without friends.’

draug-iem friend-pl.dat

We see that other prepositions, according to the traditional account, take different case values in the singular and plural. In fact all prepositions take the dative in the plural. We wish to maintain a simple rule of government, rather than having government dependent on number, which would be an odd situation. If we do this, and we also retain the morphological features and values which are fully parallel to the morphosyntactic ones, we need two additional case values. Neither of these additional case values has a unique form (they are nonautonomous). An alternative, explored within the Network Morphology framework in Corbett (2010), is to map the additional feature values onto those in (4) within the morphology, so that the morphological feature set is the simpler one. The realization of the morphosyntactic values is mediated through the lexemic hierarchy, which locates the issue appropriately at the syntax–morphology interface.

46

greville g. corbett

The particular analysis is not the issue here; rather we should consider more of these challenging systems, where the morphosyntactic requirements appear to be very different to the pattern of the morphology, and do so with an open mind as to what features and values are required within the morphology.

3.7 Conclusion

.............................................................................................................................................................................

The features which determine morphological paradigms are the morphosyntactic features. This is natural because inflectional morphology realizes the specification provided by these features. In addition we find the purely morphological features which express inflectional class. A great deal of the interest of inflectional morphology arises when the morphological forms available do not match the morphosyntactic requirements in a straightforward way. This is where we find interesting phenomena such as syncretism, defectiveness, and deponency. In some challenging examples the mismatch appears particularly severe, which raises the possibility that we should allow for less direct mapping of morphosyntactic specifications onto morphological paradigms.

Acknowledgements

.............................................................................................................................................................................

The support of the AHRC (grant AH/K003194/1) is gratefully acknowledged.

chapter 4 ........................................................................................................

INFLECTIONAL EXPONENCE ........................................................................................................

jochen trommer and eva zimmermann

4.1 Introduction

.............................................................................................................................................................................

Inflectional exponence may be understood declaratively as a systematic phonological relation between inflected word forms sharing a set of inflectional features and their bases, or derivationally as operations on base forms exposing these features. In this chapter, we discuss the types of exponence and approaches to the phenomenon which have played a prominent role in the linguistic literature of the last decades, giving an overview of the range of empirical phenomena involved, and its current theoretical understanding. A claim which is implicit in the choice of data we make is that there is no significant categorial difference between inflectional exponence and other types of exponence: Every relevant exponence pattern described for derivational (or evaluative) morphology is in principle also found in inflection. The chapter is structured as follows: Section 4.2 gives a descriptive morphologically oriented overview on types of inflectional exponence which is loosely based on the classical dichotomy between ‘Item-and-Arrangement’ versus ‘Item-and-Process’ approaches of Structural linguistics (Hockett 1954 based on the seminal work of Harris 1942), but integrates the notion of templatic exponence which has since that time become more and more important in the theoretical discussion. Section 4.3 discusses the central theoretical frameworks which have tried to reduce (or relate) non-additive exponence to phonological processes and principles. In Section 4.4, we provide a critical evaluation of approaches to linear order in exponence. There are two non-conventional features in our approach to inflectional exponence. First, we will mostly avoid the term ‘Non-concatenative Morphology’ because it conflates two, in principle, independent departures from standard prefixation and suffixation: exponence which makes no (or non-additive) use of fixed amounts

48

jochen trommer and eva zimmermann

of phonological material (discussed in Section 4.3), and non-canonical linearization of exponence (Section 4.4). Second, we will focus almost exclusively on theoretical approaches which are at home in Generative Phonology (Chomsky and Halle 1968), and its offsprings (Autosegmental Phonology, Prosodic Morphology (PM), and Optimality Theory (OT)). The reason for this choice is that morphological theories of the last decades have had very little to say about the nature of possible exponence rules (or the phonological structure of exponents) which goes beyond the positions developed in the context of American Structuralism (e.g. Matthews 1972) or cannot be found in more elaborate form in the phonological literature.

4.2 Types of Inflectional Exponence

.............................................................................................................................................................................

There are three central types of inflectional exponence discussed in the morphological literature: (1) Types of inflectional exponence 1. Additive exponence: An inflectional category is expressed by the addition of an invariant piece of phonological material to the phonological material of a lexeme (e.g. English plural through the addition of -s : lamp ⇒ lamps, tear ⇒ tears ). 2. Transformational exponence: An inflectional category is expressed by subjecting the phonological material of a lexeme to specific phonological transformations (e.g. the Hausa imperative which is formed by overwriting the underlying tone of a verb by the melody Low–High: kwáːná ⇒ kwàːná ‘spend the night’; táːʃ-ì ⇒ tàːʃ í ‘get up’) 3. Templatic Exponence: An inflectional category is expressed by adjusting the phonological material of a lexeme to a fixed shape (e.g. the plural of one class of German nouns formed by adjusting the shape of the singular to a bisyllabic foot: Kind ⇒ Kinder ‘child’, vs. Mieder ⇒ Mieder ‘bodice’). In the following subsections, we discuss examples for all three types. The relations between the different types of inflectional exponence and the interesting difficulties in assigning a certain exponence phenomenon to a specific type are discussed in detail in Sections 4.2.4 and 4.3.1.

4.2.1 Additive Exponence A simple example of additive exponence is found in Hungarian case inflection which expresses a variety of spatial (and other) concepts as illustrated in (2):

inflectional exponence (2)

49

Hungarian case suffixes base haːz ‘house’ bolt ‘shop’ kuːt ‘well’

inessive haːz-ban ‘in a house’ bolt-ban ‘in a shop’ kuːt-ban ‘in a well’

sublative haːz-ra ‘onto a house’ bolt-ra ‘onto a shop’ kuːt-ra ‘onto a well’

ablative haːz-toːl ‘from a house’ bolt-toːl ‘from a shop’ kuːt-toːl ‘from a well’

Every case category in (2) is expressed by adding an invariant string of phonological segments to the base form (e.g. ban for the inessive category). Moreover also the base nouns (e.g. haːz ) are invariantly contained in the resulting forms satisfying the condition that additive exponence involves the addition to (hence not the modification) the phonological material of the base. Both types of invariance are often obscured by phonological alternations. Thus Hungarian has a productive process of vowel harmony which fronts back vowels of suffixes after stems with front vowels as shown in (3), where -ban shows up as -ben, -ra as -re, and -toːl as -töːl : (3)

Hungarian case suffixes after back vowel stems base kert ‘garden’ öːʃ ‘ancestor’ viːz ‘water’

inessive kert-ben ‘in a garden’ öːʃ-ben ‘in an ancestor’ viːz-ben ‘in water’

sublative kert-re ‘onto a garden’ öːʃ-re ‘onto an ancestor’ viːz-re ‘onto water’

ablative kert-töːl ‘from a garden’ öːʃ-töːl ‘from an ancestor’ viːz-töːl ‘from water’

On the surface, the material added to base forms in the inessive is sometimes ban, sometimes ben, hence at this level of representation, different portions of phonological material are employed for exponence and the inessive does not instantiate solely additive exponence. However, since we are concerned with the morphophonological, not the strictly phonological side of exponence, it is convenient to abstract away from this type of complication by considering only the phonological form before the application of phonological processes, where we may assume that all instances of the inessive have ban which is then transformed by a general phonological process to ben after front-vowel bases. Obviously this strategy also extends to cases where phonological alternations modify the base of an inflected word form, as in Germanic vowel fronting (‘umlaut’), a phonological process that changes the back front vowel a to e before a syllable containing i. All Germanic languages (except Gothic) show this type of vowel harmony that is attested from about the eighth century. The examples in (4) are from Old High German and illustrate how umlaut leads to a modification of the base noun in forms with the plural suffix -i or -ir.

50 (4)

jochen trommer and eva zimmermann Old High German vowel fronting sg gast kraft aphul lamb

‘guest’ ‘strength’ ‘apple’ ‘lamb’

pl gest-i kreft-i ephil-i lemb-ir

(Mettke 1983; Paul 1989)

‘guests’ ‘strengths’ ‘apples’ ‘lambs’

On the surface, gest-i is non-additive exponence since it is not the sum of adding i to gast. But for our purposes, the morphophonological representation can be safely assumed to be gast-i which is straightforwardly additive. A second complication from which we will abstract away here is morphologically triggered allomorphy, which leads by definition to patterns of variance, but for reasons not strictly related to the nature of morphophonological inflectional exponence which is our main concern here. Additive inflectional exponence is in principle hard to distinguish from the syntactic combination of different words. Thus Hungarian also has a class of postpositions illustrated in (5) which perform similar functions as case markers and appear in exactly the same position as case markers immediately after the head noun of a noun phrase: (5)

Hungarian postpositions haːz ‘house’ kert ‘garden’

‘besides’ haːz mellet ‘besides a house’ kert mellet ‘besides a garden’

‘before’ haːz elött ‘before a house’ kert elött ‘before a garden’

‘behind’ haːz mögött ‘behind a house’ kert mögött ‘behind a garden’

To be sure, there are criteria which allow the distinction between both categories for the Hungarian case and more generally—thus postpositions do not undergo vowel harmony and behave differently with respect to ellipsis constructions and some word formation processes (see Asbury 2008; Trommer 2008a, for discussion). The phonological material which is added in additive exponence is not necessarily composed of segments, but might also consist of suprasegmental or subsegmental material. Thus the Niger-Congo language Jamsay spoken in Mali forms locative forms of nouns by adding a low tone (Heath 2008; Bye and Svenonius 2012). The examples in (6) show that the base and its tonal melody remain intact and an additional L is simply added at the end of the tonal melody of a base noun, resulting in falling tonal melodies or contour tones. (6)

Tonal locative in Jamsay bare stem a. /kaː, H/ b. /uro, H/

káː úró

‘mouth’ ‘house’

(Heath 2008; Bye and Svenonius 2012) locative /kaː, H, L/ /uro, H, L/

kâː úrò

‘in the mouth’ ‘in the house’

inflectional exponence c. /gɔː, LH/ gɔˇ ː ‘granary’ d. /numo, LH/ nùmó ‘hand’

51

/gɔː, LH, L/ gɔˇ`ː ‘in the granary’ /numo, LH, L/ nùmôː ‘in the hand’

Again the addition of non-segmental material has close parallels in syntax. Thus in intonations (question, assertion) entire sentences are combined with tonal melodies. Another instance of additive exponence where suprasegmental material is added is morphological lengthening. A famous example for morphological lengthening comes from Alabama (Montler and Hardy 1988), an Eastern Muskogean language spoken in Texas. The imperfective aspect for verbs is expressed in Alabama through geminating the onset of the penultimate syllable (7a) or through lengthening of the penultimate vowel if the antepenultimate syllable is closed or the perfective base is disyllabic (7b). (7)

Lengthening in Alabama a.

b.

perfective a.ta.kaː.li a.fi.nap.li ho.co.ba co.ba cam.po.li hof.na

imperfective a.tak.kaː.li a-fin.nap.li hoc.co.ba coː.ba cam.poː.li hoːf.na

(Samek-Lodovici 1992: 9) ‘hang one object’ ‘lock up’ ‘big (pl)’ ‘big (sg)’ ‘taste good’ ‘smell’

The characterization of lengthening as additive exponence requires some level of abstraction since the phonological element that is added to the base is only invariant at an abstract theoretical level: gemination and vowel lengthening both realize an additional mora. Lengthening could also be characterized as transformational exponence if one takes length as an intricate feature specifying segments. Under this assumption, the feature value for [±length] would be replaced and the base material would consequently be modified.

4.2.2 Transformational Exponence In contrast to additive exponence, transformational exponence either involves an inherent modification of base material or does not add an invariant portion of phonological material. A classical example of transformational exponence is umlaut (apophony) in German. Thus, for example, a number of noun plurals are expressed exclusively by fronting non-front-vowels, as shown in (8). Recall from (4) that the umlaut was phonologically triggered in an earlier stage of language development. But after weakening and losing the (-i ) affixes that triggered the umlaut, the change of the vowel quality remains the sole exponent for plurality.

52 (8)

jochen trommer and eva zimmermann Umlaut: plural formation by vowel fronting in Modern High German singular Bruder Tochter Vater

‘brother’ ‘daughter’ ‘father’

plural Brüder Töchter Väter

‘brothers’ ‘daughters’ ‘fathers’

To be sure, German umlaut may be partially understood as additive according to the definition of Additive Exponence in (1): it involves an invariant amount of phonological representation (the phonological feature [–back]). Where umlaut becomes non-additive is the modification of the base which is involved by this step: the [+back] specification of the base vowel is altered, deleted, or ‘overwritten’. Similar examples for featural overwriting also abound for consonants, where they are usually referred to under the term mutation). A famous example can be found in the Bantu language Aka, where the class 5 singular is marked through voicing the initial consonant as shown in (9) (Akinlabi 1996). The plural forms have a segmental prefix, but show the underlying contrast between root-initial voiced and voiceless obstruents, which is neutralized by the mutation process in the singular. Again, an invariant portion of phonological material [+voice] is added to the base, but the [–voice] specification of the base itself remain unrealized, hence is overwritten: (9)

Consonant mutation in Aka a.

b.

singular, class 5 dèŋgé gásá bàpùlàkà gìnì b`ɔkí bèlèlè gòàlà

plural, class 6 mà-tèŋgé mà-kásá mà-pàpùlàkà mà-kìnì mà-p`ɔkí mà-bèlèlè mà-gòàlà

(Akinlabi 1996: 286) ‘piercing tool’ ‘palm branch’ ‘lung’ ‘fly’ ‘arch of the eyebrow’ ‘sound of a waterfall’ ‘game of imitation’

Overwriting of base specifications with the phonological material of a morphological exponent is not only attested for segmental specifications but for suprasegmental features as well. An example is tonal overwriting, illustrated in (10) with data from Hausa, a Chadic language spoken in large parts of western Africa. The imperative in Hausa is not marked through adding the fixed tonal melody LH to the base melody but through replacing any underlying base tones with this fixed tone sequence. (10)

Tonal overwriting in the Hausa imperative a. b. c.

underlying form H kwáːná HL táːʃ ì HLH káràntáː

imperative LH kwàːná LH tàːʃí LH kàràntáː

(Newman 2000: 262–3) ‘spend the night’ ‘get up’ ‘read’

inflectional exponence

53

Examples for exponents that leave their base intact, but add a variable portion of phonological material exist as well, namely reduplication. The example in (11) illustrates how full reduplication in the plural of Indonesian nouns leaves its phonological base intact, but adds phonological material to the base that is not invariant since it is a function of the base material. In other words, the shape of the reduplicant can only be determined by knowing (inspecting) the shape of the base, rendering this pattern non-additive according to the definition in (1). (11)

Total reduplication in Indonesian (Sneddon 1996; Aronoff and Fudeman 2010) kuda-kuda rumah-rumah singkatan-singkatan perubahan-perubahan

‘horses’ ‘houses’ ‘abbreviations’ ‘changes’

An interesting borderline case is partial reduplication. For example in the Austronesian language Mokilese, the progressive aspect of verbs is expressed through prefixing a reduplicant to the stem that has the size of a bimoraic syllable (Blevins 1996; McCarthy and Prince 1996). The base remains unmodified and the phonological material added to the base has an invariant size: one syllable. But although the size is invariant, the segmental material that is affixed is once again variable since it is a function of its base. Partial reduplication is therefore another instance of transformational exponence. (12)

Partial reduplikation in Mokilese bare stem pɔdɔk mwiŋe kasɔ sɔːrɔk tʃaːk

‘plant’ ‘eat’ ‘throw’ ‘tear’ ‘bend’

progressive pɔd- pɔdɔk mwiŋ- mwiŋe kas- kasɔ sɔː-sɔːrɔk tʃaː- tʃaːk

(Harrison 1976; Blevins 1996) ‘planting’ ‘eating’ ‘throwing’ ‘tearing’ ‘bending’

In fact there are types of transformational exponence which neither add a fixed amount of phonological structure nor leave their base intact, hence are deeply non-additive. In subtractive morphology, exponence is achieved by removing a fixed amount of phonological material. The best-established case is Tohono O’odham, an Uto-Aztecan language spoken in Arizona and Mexico. The perfective in Tohono O’odham is expressed by removing the final segment of the verb stem (13):

54 (13)

jochen trommer and eva zimmermann Subtractive morphology in Tohono O’odham (Fitzgerald and Fountain 1995: 5–6) imperfective perfective máːk máː ‘giving’ híːnk híːn ‘barking’ híhim híhi ‘walking’ (pl) gátwid gátwi ‘to shoot object’ híkˇck híkˇc ‘cutting’ cˇ ɨʔɨwid cˇ ɨʔɨwi ‘covering’ nákog náko ‘enduring’

Similarly, morphological shortening does not add any constant portion of phonological information to the base and modifies the base by making a long base vowel short to express a specific inflectional category. An example can be found in the Eastern Franconian dialect of Upper German spoken in the Taubergrund area, where the exponence of plural for a subclass of nouns consists in vowel shortening (14):1 (14)

Shortening in Taubergrund German singular riːs fiːʃ ʃniːds fleːk

plural ris fiʃ ʃnids flek

(Seiler 2008: 1, citing Heilig 1898)

‘crack’ ‘fish’ ‘cut’ ‘blot’

A further type of deeply non-additive transformational exponence is metathesis— reordering of segments. Studies on metathesis mainly concentrate on the questions whether the reordering of segments is a truly productive synchronic phonological process in a language or what (phonetic) cues influence the likelihood of its appearance (cf. e.g. Hume 2001 or 2004 for discussion and references). That the reordering of segments expresses an inflectional category has been argued for only a handful of languages (Stonham 2006): for example in the Austronesian language Rotuman (Churchward 1940; McCarthy 2000), the Penutian languages of Sierra Miwok (Freeland 1951; Broadbent 1964; Sloan 1991), and some Salishan languages (e.g. Clallam: Thompson and Thompson 1971; Saanich: Montler 1986; Stonham 1994; or Upriver Halkomelem: Galloway 1993). In addition, it is a striking observation that metathesis is never the only exponent of a morpheme. Virtually all cases where metathesis seems to be a productive means to express some morphological category are additive in the sense that metathesis is a side effect of adding weight to the underlying 1 In fact, under the moraic analysis of vowel length (Hayes 1989; Odden 2011), vowel shortening is a special case of subtractive morphology since one of the underlying moras of the long vowel must be removed to produce a short vowel.

inflectional exponence

55

base. This is discussed in some more detail in Section 4.3.2. Example (15) illustrates morphological metathesis from Saanich, a Central Salishan language from British Columbia. Here the continuative aspect of verbs is marked by switching the order of the second root consonant and the preceding vowel.2 (15)

Metathesis in Saanich non-continuative q’p’ət ´ sxət ´ t’sət ´

(Montler 1986; Kurisu 2001) continuative q’əp’t ´ səxt ´ t’əst ´

‘patch’ ‘push’ ‘break’

4.2.3 Templatic Exponence In contrast to additive and transformational exponence, templatic exponence is not (or not exhaustively) defined with respect to the change(s) a given base undergoes, but employs an output target T, where bases are adjusted by different means to meet the phonological conditions of T. A case in point is the formation of the habilitative in Cupeño, an extinct Uto-Aztecan language that was formerly spoken in Southern California. As can be seen in (16), the habilitative is formed by ensuring that the part of the output which follows the stressed vowel has exactly the size of two syllables. By comparing the forms with their nonhabilitative counterparts, it becomes clear that the stress of the base is kept and that the templatic restriction is achieved through copying the final base vowel as often as necessary together with insertion of epenthetic [ʔ] (McCarthy and Prince 1990; Crowhurst 1994). Consequently, no change is necessary to form the habilitative for bases with two (or more) post-stressed syllables (16a). But one or two copies of the final vowel are inserted whenever the stressed syllable is on the penultimate or ultimate syllable in the base (16b) and (16c). (16)

Templates in Cupeño a.

stem (pí)(nəʔwəx) (xá)(ləyəw)

(McCarthy and Prince 1990; Crowhurst 1994) habilitative (pí)(nəʔwəx) (xá)(ləyəw)

‘sing enemy songs’ ‘fall’

' ‘see’ non-continuative ∼ continuative), For stems containing a sonorant (e.g. k’wə´ nət ∼ k’w´ənt metathesis is not surface apparent due to a general phonological process that inserts a [ə] between sonorants and preceding obstruents. A natural formal implementation of this fact is that morphological metathesis precedes [ə]-insertion. Thus the non-continuative of ‘see’ would be underlyingly k’wnət, ' ). [ə]-insertion then targets the non-continuative which results by metathesis in the continuative k’w´ənt because this still has a sonorant right-adjacent to an obstruent (k’wnət ⇒ k’w´ənət ), but not the con' , where morphological metathesis has already separated the obstruent and the following tinuative k’w´ənt sonorant. See Leonard and Turner (2010) for the full range of data and for critical discussion of the metathesis analysis based on [ə]-insertion. 2

56

jochen trommer and eva zimmermann b.

c.

(pá)tʃik (nə)nəŋ ´ (yúy)muk (ŋáŋ) (nəŋán) (həlyəp) ´

(pá)(tʃiʔik) (nə)nəʔəŋ) ´ (yúy)(muʔuk) (ŋá)(ʔaʔaŋ) (nəŋá)(ʔaʔan) (həlyə)(ʔəʔəp) ´

‘leach acorns’ ‘play pion’ ‘be cold’ ‘cry’ ‘take’ ‘hiccup’

Another example for templatic exponence can be found in Sierra Miwok3 (Freeland 1951; Broadbent 1964; Sloan 1991) where the whole base is adjusted to specific templatic requirements. This pattern therefore differs crucially from the Cupeño Habilitative, where only part of the base (the portion following the stressed vowel) is subject to the templatic restriction to two syllables. Examples for four different templatic exponents in Sierra Miwok are given in (17). At first glance, the templates in Sierra Miwok are part of (exponents of) the suffix since the choice for one or the other stem form depends on the following suffix, for example the verbalizer suffix -j must always be preceded by a stem of the form CVCVː (Broadbent 1964: 58). But the templates in Sierra Miwok must be ‘regarded as in some measure having grammatical value in itself, and not being entirely accompanied by suffixation’ (Freeland 1951: 96). This is especially apparent since some stems serve as a form of the verb without any suffix (stem 4 = infinitive). In addition, every stem is bound to a specific (core) meaning, for example the primary stem to present tense or the second stem to future and past tense. We can therefore conclude that the templatic restrictions are exponents of their own. (17)

Four stems in Central Sierra Miwok Stem 1 a. b. c. d. e. f. g. h.

tujáːŋ poláːŋ wɨ´kt̪ɨ t̪úyku jɨ´nːa líwːa mɨ´ːŋ háːt

(CVCVːC)

´ (CVCCV)

´ (CVCːV)

´ (CVːC)

Stem 2 CVCː/CVCVC tujáŋː poláŋː wɨkɨ´t̪ː t̪úyukː jɨnáʔː liwaʔː mɨ´ŋː hátː

(Freeland 1951: 94–5) Stem 3 CVCːVC tújːaŋ pólːaŋ wɨ´kːɨt̪ t̪úyːuk jɨ´nːaʔː líwːaʔ mɨ´ŋːɨʔ hátːïʔ

Stem 4 CVCCV tújŋa pólŋa wɨ´kt̪ɨ t̪úyku jɨ´nʔa líwʔa mɨ´ŋʔɨ hátʔï

‘jump’ ‘fall’ ‘burn’ ‘to poison’ ‘kill’ ‘speak’ ‘swim’ ‘step on’

3 There are three main dialects of Sierra Miwok: Southern, Northern, and Central Sierra Miwok. The examples here are taken from the Central dialect but the same generalizations about templates apply to at least the Southern dialect as well, which is described and discussed by Broadbent (1964) and Sloan (1991).

inflectional exponence

57

4.2.4 The Relation and Interaction between the Different Types of Exponence Ontological relations between different exponent types: Additive exponence as well as templatic exponence might be understood as special cases of transformational exponence. Obviously, adding phonological material is a kind of phonological operation/transformation, whereas templatic morphology typically involves the application of different transformations in different contexts to ensure that the form conforms to a templatic shape. Reconsider for example the templatic exponence in Central Sierra Miwok from (17). The phonological transformations listed in (18) apply to adjust bases of different phonological forms to the required templatic shape. On their own, these phonological operations can be interpreted as transformational (or even additive) exponence but it is clear that the combination and interaction of all these operations serve to conform to a templatic requirement. This is apparent since hardly any of these operations ever applies in isolation; rather there are always two or more non-concatenative strategies that apply. (18)

Strategies to adjust bases to templatic requirements in Central Sierra Miwok example lengthening epenthesis metathesis shortening stress shift

háːt líwːa wɨ´kt̪ɨ tujáːŋ poláːŋ

(stem 1) (stem 1) (stem 1) (stem 1) (stem 1)

hátː líwːaʔ wɨkɨ´t̪ tujáŋː pólːaŋ

(stem 2) (stem 3) (stem 2) (stem 2) (stem 3)

However, borderline cases between transformational and templatic exponence abound. Thus, German past participles are formed (apart from suffixation) by prefixing the syllable ge to the verb if this starts with a stressed syllable (e.g. kauf ⇒ ge-kauf-t ‘buy’, píckel ⇒ ge-píckel-t ‘pickle (in leather manufacture)’), otherwise if the base starts with an unstressed syllable only suffixation applies (trompét ⇒ trompétet ‘trumpet’, ratifizíer ⇒ ratifizíer-t ‘ratify’). This may be interpreted as a partial template requiring participles to start with an unstressed syllable, but also as a straightforward transformation. Similarly, most cases of overwriting may be understood as templatic since they ensure that output forms have a (partially) fixed phonological shape. Thus the Hausa imperative is a tonal template similar to the prosodic template found in Cupeño and Miwok. More generally, we may understand Additive Exponence as the prototypical form of an exponence operation combining two canonical properties: First it adds a fixed amount of phonological material to the base, or in more declarative terms: exponence for a given category C is additive if all forms of category C share a characteristic piece of phonological representation (the Fixed Target Property). Second, Additive Exponence leaves its base intact—the phonological representation of every word form exhibiting

58

jochen trommer and eva zimmermann

Additive Exponence is an extension of the phonological representation of its base (the Base Extension Property). Deeply transformational exponence such as subtractive morphology and metathesis have neither of these properties. Overwriting and reduplication take a somewhat intermediate position in exhibiting one of these properties, but not the other one. This is shown schematically in (19): (19)

Cross-classification of exponence by canonical properties Base Extension Property

Fixed Target Property

✓

✓

✗

✓

✓

✗

Additive Exponence Umlaut Mutation Tonal overwriting Reduplication Subtraction Shortening Metathesis Stress shifting

✗

✗

Functional co-occurrence of different exponence types: Virtually every type of transformational and templatic exponence co-occurs with additive exponence to achieve realization of a single morphosyntactic category, a situation which might be called ‘symbiotic’. Thus in contrast to pure tonal overwriting in the imperative, Hausa expresses the ventive category of verbs by tonal overwriting and affixation of -o : (20)

Tonal overwriting and affixation in the Hausa ventive a. b. c. d.

underlying form LH fìtáː HL fáɗì HLH gángàráː LHL tàimákàː

H H H H

ventive fít-óː fáɗ-óː gángár-óː táimák-óː

(Newman 2000: 663)

‘go out’ ‘fall down’ ‘roll down’ ‘help’

Another example is the frequent co-ocurrence of morphological lengthening with suffixation. Many cases of morphological lengthening cited in the literature are similar to the Zuni data in (21):4 a segmental suffix attaches to the right edge of the base and in addition, the preceding (stem) vowel is lengthened.

4 Other examples are morphological lengthening in Aymara (Kim 2003), Arbizu Basque (Sprague 2004), or Quechua (Weber and Landerman 1985).

inflectional exponence (21)

Affixation and lengthening in the Zuni plural lupa homata

‘box of ashes’ ‘juniper tree’

lupaːweʔ homataːweʔ

59

(Kirchner 2007: 11)

‘boxes of ashes’ ‘juniper trees’

Functional co-occurrence raises the question how tight the relation between affixation and transformations (or templates) is generally and in specific cases. Consider for example the co-occurrence of German umlaut with different inflectional affixes. There are inflectional categories such as the 1sg of verbs which never co-occur with umlaut (22d), whereas plural formation by umlaut occurs without segmental affixation (22b) and plural formation by segmental affixation without umlaut (22c). Finally there are categories such as the 2sg which exhibit umlaut with many, but not all stems (22a), and similarly there are nouns with Ø-plural which do not umlaut (22b) (schlagen, ‘to beat’; zagen, ‘to worry’; Vater, ‘father’, Oma, ‘grandma’; Haken, ‘hook’): (22)

Co-occurrence of umlaut and affixation in German a. b. c. d.

2sg pl pl 1sg

affix -st -Ø -s -e

umlaut schlagen schläg-st Vater Väter-Ø – –

no umlaut zagen zag-st Haken Haken-Ø Oma Oma-s schlagen schlag-e

The crucial point here is that there is no strictly biunique conditioning between umlaut and co-occurring segmental affixes. The only valid generalization we can draw is that umlaut will not occur if there is a 1sg affix (or another ‘umlaut-less’ affix). We might consider the relation of transformational exponence and affixation on a par with the co-occurrence of segments (e.g. s and t in German 2sg -st ) in a segmental affix. The other extreme is to see the combination of two exponence patterns as cases of multiple exponence comparable to cases where the same (or overlapping) categories are expressed by different segmental affixes. An example can be found in the Eastern Cushitic language Jubba Maay. The marking of plural involves allomorphy as can be seen in (23). Plural for vowel-final nouns is marked by the suffix -yal. Consonant-final nouns, on the other hand, optionally take the suffix -o, the suffix -yal, or even both suffixes to mark their plural. The latter case is an instance of multiple exponence since -o and -yal both mark plural on their own. (23)

Multiple exponence in Jubba Maay a.

singular plural Vowel-final nouns šati šati-yal buundo buundo-yal aweesa aweesa-yal

(Paster 2006)

‘shirt’ ‘bridge’ ‘worn’

60

jochen trommer and eva zimmermann b.

Consonant-final nouns mukulal mukulal-o eey eey-o af af-o

∼ mukulal-yal ∼ eey-yal ∼ af-yal

∼ mukulal-o-yal ∼ eey-o-yal ∼ af-o-yal

‘cat’ ‘dog’ ‘mouth’

Even subtractive morphology may be in a functional unit with additive affixation. For example in Yine, a Maipurean language spoken mainly in the Peruvian Amazon, certain suffixes are always accompanied by the deletion of the preceding vowel as can be seen in (24) where suffixes triggering deletion are underlined (Matteson 1965; Lin 1997a, 1997b, 2005; Pater 2007; Zimmermann 2013). (24)

Vowel deletion in Yine neta -lu walapu -ni çema -çe –ta homkahita -ka heneka -sa

(Matteson 1965) netlu walapni çemçeta homkahitka heneksa

‘I see him’ ‘the past summer’ ‘have never, never heard’ ‘be followed’ ‘distribute’

On the surface, no phonological property distinguishes these suffixes from suffixes that do not trigger vowel deletion in Yine. This is apparent since pairs of homophonous suffixes exist where one triggers deletion and the other does not. Examples for this are given in (25). The data in (25a) are particularly striking since not only are the two suffixes homonymous but the two stems to which the different affixes attach are phonologically quite similar as well. Only the nominalizing suffix -nu triggers deletion, the homophonous anticipatory suffix does not. (25)

Homophonous suffixes in Yine a.

b.

(Matteson 1965; Pater 2010)

i.

hata -nu shine-nmlz

hatnu

‘light, shining’

ii.

heta -nu see-ant

hetanu

‘going to see’

i.

homkahita -ka follow-pass

homkahitka

‘to be followed’

ii.

tçirika -ka rub-smlf

tçirikaka

‘to ignite’

Finally, there are also cases where a single category is expressed by different nonadditive operations.Thus the Anywa Antipassive not only shortens base vowels (púːr ⇒ púr, ‘cultivate’), but also shifts [–ATR] vowels to [+ATR] (càm ⇒ c ʌ` m, ‘eat’), hardens base-final liquids (k ʌ` l ⇒ k ʌ` t, ‘take away’), and effects various tonal changes

inflectional exponence

61

(múl ⇒ mùt, ‘touch’, Reh 1993: 222–6). The situation in related Western Nilotic languages is similar, cf. the Shilluk noun kâːl, ‘cattle camp’ which forms its plural k ʌ` ¯ːːt by stem-final consonant hardening, changing its falling tone to mid, raising [a] to its [+ATR] counterpart [ʌ], and further lengthening of its long vowel (Remijsen this volume: Chapter 24).

4.3 Non-additive Exponence as Phonology

.............................................................................................................................................................................

A central goal of linguistic research on exponence in the last decades has been to reduce non-additive exponence as far as possible to phonological processes and/or principles. There are two main lines of research, one which assumes that non-additive exponence is the consequence of additive operations which affix non-segmental (i.e. suprasegmental, subsegmental, or defective segmental) phonological material. Following Bermúdez-Otero (2012), we will call this approach Generalized Nonlinear Affixation, and the hypothesis that all non-additive exponence is derived in this way the Concatenativist Hypothesis (Section 4.3.2). The second line of research pursues the idea that non-additive morphology is the consequence of phonological rules or constraints which are restricted to specific morphological contexts (Section 4.3.1). Of course the two approaches are logically not mutually exclusive.

4.3.1 Non-additive Exponence as a Phonological Constructicon Exponence as emergence of unmarked phonology: Consider again German noun plurals which instantiate a kind of template since virtually every noun of canonical phonological shape ends in the plural in a bisyllabic trochee. Either the base is already of this shape, then the plural is prosodically identical to the base (the nominative singular (26a)), or, on the other hand, if the base ends in a stressed vowel, [ə] or [ɐ] is suffixed to it (26b) ([ɐ] is rendered in German orthography by , [ə] by ): (26)

The bisyllabic trochee template in German a. b.

sg (Sú.cher) (Sé.gel) (Líed) (Schúh)

pl (Sú.cher) (Sé.gel) (Líe.der) (Schú.he)

‘searcher’ ‘sail’ ‘song’ ‘shoe’

In the optimality-theoretic analysis of Wiese (2009), this template is captured by assuming that there are general well-formedness constraints (abbreviated in (27) and (28) by the cover term ‘Trochee’) which specify a binary syllabic trochee as the optimal

62

jochen trommer and eva zimmermann

prosodic shape. The Trochee constraints are ranked as a block above the constraint Dep Segpl that penalizes segment epenthesis in plural forms, but below Dep Segsg , which serves the same function in singular forms. The net effect is that the Trochee constraints are satisfied by schwa-epenthesis in plural (27b), but not in singular forms (27a): (27)

Singular and plural of a monosyllabic noun Input: Túch ☞

☞

i. (Túch) ii. (Tú.che)

Dep Segsg *!

Trochee * *!

i. (Túch) ii. (Tú.che)

Dep Segpl a. singular *

b. plural

In bisyllabic nouns, epenthesis does neither take place in singular nor in plural forms, since the unmodified base already perfectly satisfies Trochee: (28)

Singular and plural of a bisyllabic noun Input: Ségel ☞ ☞

i. (Sé.gel) ii. (Sé.ge).le i. (Sé.gel) ii. (Sé.ge).le

Dep Segsg

Trochee

Dep Segpl a. singular

*! *!

b. plural

Thus under this analysis there is neither a plural morpheme nor a plural template. The templatic shape of the plural simply emerges from general phonological wellformedness constraints, and the fact that some constraints (in this case Dep Seg) may be indexed with respect to the morphological context in which they apply. The general approach of phonological sensitivity to morphological information exemplified by Wiese’s analysis is known in the OT-literature as indexed constraints. A slightly different approach is Cophonology Theory advocated most forcefully by Inkelas and Zoll (2005), where the constraints themselves are ‘blind’ to morphological features, but different morphological constructions may be associated with different constraint rankings. Both approaches can also be applied to non-additive exponence which is non-templatic/deeply transformational. Thus the pattern of metathesis in Saanich we have cited in (15), where the second C and the first V are reordered in non-continuative forms for bases with complex onsets (e.g. q’p’ ´ət ‘patch’ ⇒ q’ ´əp’t ‘patching’) could be captured as in (29). Metathesis obtains in continuative forms to avoid violations of complex onsets since *ComplexOnset is higher ranked than the faithfulness constraint penalizing reordering of segments, Linearitycont (30b).

inflectional exponence

63

In non-continuative forms, metathesis is blocked by undominated Linearitynon-cont (30a):5 (29)

Metathesis as emergence of the unmarked Input: sxət ´ ☞ i. sxət ´ ii. səxt ´ i. sxət ´ ☞ ii. səxt ´

Linnon-cont *ComplexOnset Lincont *!

* *!

a. non-continuative *

b. continuative

The phenomenon that phonologically unmarked structure in a language shows up only under some grammatically circumscribed circumstances is known as The Emergence of the Unmarked (TETU) in the OT-literature. A substantial body of research on prosodic morphology over the last two decades especially on reduplication and truncation has shown that Emergence of the Unmarked accounts for a wide range of cross-linguistic variation in exponence (McCarthy and Prince 1994, 1995; Downing 2006). Now, assuming that morphologically sensitive phonology is the only source of exponence would lead to the prediction that all exponence is phonologically optimizing with respect to underlying bases, a claim which is obviously too strong. Consider for example consonant mutation in the Nilotic language Nuer. Negative present participle mutation (30b) fulfils the Emergence of the Unmarked prediction since it makes obstruents non-continuant (i.e. stops) and voiceless, arguably the unmarked incarnations of obstruents. However, past participle formation effects just the reverse change for the feature [±continuant], turning stops into fricatives (30c). Even if both morphological constructions may differ in constraint ranking, it seems 5 That ‘metathesis’ in Saanich is the side effect of morphologically sensitive phonological processes is also argued for by Leonard and Turner (2010). In their recent reanalysis of the Saanich continuative, they maintain that the process in question is in fact not even metathesis, but rather morphologically contrastive [ə]-insertion. Their crucial observation is that all verb forms undergoing ‘metathesis’ contain a [ə] (i.e. ‘metathesis’ never affects any vowel other than [ə]) and that occurrences of [ə] in Saanich are phonologically predictable in all other contexts (Kinkade 1998; Shaw et al. 1999): they result either from the reduction of full vowels or from different rules of vowel epenthesis in consonant clusters (see e.g. fn.3). However, this type of analysis faces the dilemma that the distinctive positioning of schwa in continuative and non-continuative forms is by its very nature not phonologically predictable. In the context of the current discussion, this leads to the more specific problem that the position of [ə] in the non-continuative cannot be predicted by a phonological constraint (whether indexed or not) given that the resulting form (e.g. sxə´t ) is phonologically more marked than the competing form with an interconsonantal [ə] (sə´xt ). Leonard and Turner (2010) propose to solve this Gordian knot by stipulating that all [ə]’s that are not predicted by the general demand to break up consonant clusters, are inserted into different positions in order to keep continuative and non-continuative forms ‘clearly recognizable’ (Leonard and Turner 2010: 7), i.e. distinguishable. Whereas they don’t work out this point in technical detail, this solution seems to be close in spirit to Antifaithfulness and Realize Morpheme constraints in OT, an approach which we discuss later in this subsection.

64

jochen trommer and eva zimmermann

unlikely that both changes (and especially the latter one) improve phonological markedness: (30)

Mutation in Nuer non-finite forms

a. Infinitive b. Neg.pres.part. c. Past part.

(Crazzolara 1933)

‘overtake’

‘hit’

‘pull out’

‘scoop hastily’

coβ còp cof

jaç jaːc jaːç

guð gut̪ guθ

kêp kɛp k`ɛf

[–voice –continuant] [–voice +continuant]

One possible reaction to markedness-increasing exponence is to marry it with affixation as a separate additive exponence device. Thus Alderete et al. (1999) assume that in cases of reduplication exhibiting segmental overwriting of the reduplicant, markedness-decreasing overwriting (cf. replacement of the onset by [ʔ] in Tübatulabal to ːyan ⇒ ʔoː-doyan, ‘he is copulating’ ʃɨʔɨwɨ ⇒ ʔɨ-ʃɨʔɨwɨ, it looks different) is phonological TETU. On the other hand, markedness-increasing overwriting as in English schm -word reduplication (e.g. table ⇒ table-schmable ) follows from segmental additive affixation. The problem with this move is that it seems to largely void the phonological approach from phonological predictiveness. If every case of nonoptimizing exponence can be reanalysed via affixation, then the claim that exponence is phonologically optimizing becomes empirically vacuous. Alternatively, many authors have assumed that markedness-increasing exponence is a consequence of constraints requiring specific phonological shapes for particular morphological constructions (Hammond 1997; Russel 1999; MacBride 2004) (cf. also the parsimonious, beautiful, and arcane formalism developed by Golston 1996 where morphemes are reconceptualized, not as constraints but as constraint violations). The most interesting offspring of this line of thinking is the hypothesis that exponence may be the consequence of constraints requiring the phonological non-identity of paradigmatically related forms, an idea which makes specific predictions about patterns of possible exponence. Antifaithfulness: One influential version of this approach, transderivational antifaithfulness (TAF) theory is developed by Alderete (2001) (see also Alderete 1999). TAF constraints require that the output of a derived form and the output of its morphological base differ for a specific phonological dimension. More specifically, Alderete assumes that for every faithfulness constraint such as Ident [voice] there is a corresponding antifaithfulness constraint (here: ¬Ident [voice]). (31)

Faithfulness and anti-faithfulness for [voice] a. b.

Ident [voice] Corresponding segments agree in the feature [voice]. ¬Ident [voice] It is not the case that corresponding segments agree in the feature [voice].

inflectional exponence

65

The tableau in (32) shows how (31b) ranked above (31a) allows the derivation of voicing polarity in Dholuo. ¬Ident [voice] requires changing the voicing of at least one segment, which rules out the c-candidates. However, additional voicing changes as in the b-candidates are blocked by Ident[voice]: Alderete uses this constraint to analyse voicing polarity in the plural formation of the Western Nilotic language Dholuo (e.g. bat ⇒ bed-e, ‘arm(s)’, kiːdí ⇒ kíːt-ê, ‘stone(s)’) as in (32): (32)

Voicing exchange in Dholuo as antifaithfulness Base

Derivative ☞

i. /bat/ ☞ ii. /cogo/

a. bed-e b. ped-e c. bet-e a. cok-e b. ɟok-e c. cog-e

¬Ident [voice]

*!

*!

Ident [voice] * **! * *!*

Alderete’s system seems to be well-suited to capture cases of deeply transformational exponence such as polarity and subtractive morphology (see Horwood 2001 for an analysis in terms of the constraint ¬Max, the antifaithfulness correspondent of the standard-OT constraint Max, which requires realization of segments in the output). Realize Morpheme: A substantially different take on deriving exponence from requirements on paradigmatic distinctness has been developed by Kurisu (2001). In contrast to Alderete who assumes that particular patterns of exponence (such as voicing polarity) correspond to constraints requiring distinctivity for specific phonological dimensions, Kurisu assumes the single but fully general constraint requiring phonological distinctivity of output forms in (33): (33)

Realize Morpheme (RM) (Kurisu 2001: 39) Let α be a morphological form, β be a morphosyntactic category, and F(α) be the phonological form from which F(α+β) is derived to express a morphosyntactic category β. Then RM is satisfied with respect to β iff F(α+β)̸=F(α) phonologically.

Since for Realize Morpheme, any phonological difference between output forms is equivalent, there must be different means to capture the fact that morphological constructions differ in the means of exponence they employ. This is achieved by indexed faithfulness constraints for specific morphological constructions. Imagine for example a language which forms inflectional category C1 by metathesis, and category C2 by segment subtraction. This could be modelled in Kurisu’s approach as follows. For C1,

66

jochen trommer and eva zimmermann

Max is ranked above Linearity. Consequently metathesis (34a) leads to less serious constraint violations than deleting a segment (33b): (34)

Inflectional category C1—metathesis /truk/ ☞

a. turk b. tru c. truk

Realize Morpheme

MaxC1

LinearityC1 *

*!

*!

On the other hand, among the faithfulness constraints specific to C2, Linearity is ranked above Max, resulting in subtraction: (35)

Inflectional category C2—subtraction /truk/ ☞

a. turk b. tru c. truk

Realize Morpheme

LinearityC2 *!

*!

MaxC2 *

The ranking Realize Morpheme ≫ {Linearity C2 , Max C1 } ≫ {Linearity C1 , Max C2 } therefore predicts a language where one morphological category C1 is marked by metathesis and another morphological category C2 is marked by subtraction. An immediate complication for Kurisu’s approach not found with Antifaithfulness are cases of exponence where non-additive exponence is obligatorily accompanied by standard affixation. Affixation alone would already be enough to guarantee phonological distinctness of output forms and any candidate undergoing an additional phonological operation is expected to be harmonically bound (see Kurisu 2001 for a solution in terms of phonological opacity). Conversely, a crucial advantage of the RM approach over Antifaithfulness is that it gives rise to a simple account of Non-concatenative Allomorphy, that is, patterns where an inflectional category C is exposed by different patterns of exponence depending on the phonological shape of the lexeme to which C is applied. Consider again continuative formation in Saanich (15). In fact, metathesis only applies to verbs which start with a complex onset (36a). Monosyllabic bases show instead reduplication (36b), and bisyllabic bases infixation of the default consonant ʔ (36c):6 (36)

Continuative allomorphy in Saanich non-continuative continuative a. q’p’ət ´ q’əp’t ´ sxət ´ səxt ´ t’sət ´ t’əst ´

(Montler 1986) ‘patch’ ‘push’ ‘break’

6 In addition to phonologically conditioned allomorphy, Saanich continuative formation also involves a certain amount of lexical idiosyncrasy. Thus Leonard and Turner (2010) argue contra Kurisu that the choice between the formation patterns in (36b) and (36c) is conditioned by idiosyncratic lexical class, not by the phonological shape of the base.

inflectional exponence b.

c.

qén qwəl ´ kwúl ʔˈit’2V wˈeqes tʃˈaq’wV

qéqən’ qwəq ´ wəl w k úkwəl ʔˈiʔt’2 wˈeʔqes tʃˈaʔq’w

67

‘steal’ ‘say’ ‘school’ ‘get dressed’ ‘yawn’ ‘sweat’

Crucially (and in contrast to Antifaithfulness) all changes in (36) are in principle equivalent for Realize Morpheme. Thus the analysis in Kurisu (2001) builds on the idea that for different phonological inputs, different phonological operations lead to a phonologically optimal output. Remaining problems: Capturing morphological exponence by phonologically sensitive phonology faces a number of substantial problems. Conceptually, allowing phonological constraints (or constraint rankings) to refer to arbitrary morphological contexts substantially undermines linguistic modularity and leads to the proliferation of possible analyses. In the words of Bermúdez-Otero (Bermúdez-Otero 2012: 79) ‘lexically specific constraints suffer from a severe lack of empirical content and heuristic power’. Conversely, phonological mechanisms are unlikely to be able to produce all types of exponence so this type of approach doesn’t provide a completely unified account of exponence. A specific problem for Antifaithfulness and Realize Morpheme is that they fail to predict that non-concatenative exponence is typically adjacent to additive exponence for the same category. Thus in Dholuo, voicing polarity in the plural occurs under strict adjacency to segmental plural affixes, excluding the possible alternative candidates *pet-e and ɟog-e where the voicing of the first consonant alternates (cf. tableaux (32)). For Dholuo this problem might be fixed by a high-ranked faithfulness constraint which specifically protects the underlying voicing specification of obstruents in root-initial syllables (an instance of positional faithfulness in the sense of Beckman 1997). However, this leads to the counterintuitive prediction that prefixes which form a morphological unit with a mutation process should exhibit an anti-adjacency effect, where mutation tends to apply to non-initial positions of bases. A concrete counterexample against this expectation is umlaut in Chamorro, where prefixes trigger vowel fronting on root-initial syllables (cf. Kaplan 2011, and references cited there). Antifaithfulness and the Realize Morpheme approach also seem to overgenerate typologically. Thus, as argued in Section 4.2.2, no language seems to have a productive pattern of unconditional morphological metathesis, that is, productive metathesis which stands not in an allomorphic relationship to other means of exponence, but this is a pattern which can be straightforwardly derived in both frameworks (cf. the tableau in (34)). Finally, the prediction that certain types of exponence systematically involve the emergence of phonological unmarkedness is an important insight, but the challenge to existent theories is to specify exactly under which conditions unmarkedness should emerge. In this context it is

68

jochen trommer and eva zimmermann

a surprising observation that the following seems to hold typologically with very few exceptions: (37)

Prosodic Primacy: Emergence of (non-strictly prosodic) unmarkedness is restricted to the context of prosodic templates

Thus there seem to be no languages where, say, plural is expressed exclusively by simplifying all complex onsets in a base to single ones, or by replacing vowels by schwa, whereas such changes are abundant in reduplicative templates. Statement (37) is unexpected in TETU-approaches to templatic exponence since prosodic templates here are simply one special case of TETU effect in principle equivalent to other phenomena of this type.

4.3.2 Generalized Nonlinear Affixation and the Concatenativist Hypothesis The Concatenativist Hypothesis (advocated under different names by Lieber 1992; Stonham 2006; Bye and Svenonius 2012) is the claim that all exponence is additive while apparently transformational and templatic patterns of exponence are the consequence of phonological alternations triggered by the addition of phonological material: (38)

The Concatenativist Hypothesis: Exponence = Additive Exponence + Phonological Alternations

Whereas the Concatenativist Hypothesis is by no means universally accepted in the linguistic community, it is the (often implicit) driving force behind some of the most important theoretical developments in work on the morphology–phonology interface, namely autosegmental morphology and prosodic morphology. Consider again the tonal locative in Jamsay: (39)

Tonal locative in Jamsay bare stem a. /kaː, H/ b. /uro, H/ c. /gɔː, LH/ d. /numo, LH/

káː úró gɔˇ ː nùmó

(Heath 2008; Bye and Svenonius 2012) ‘mouth’ ‘house’ ‘granary’ ‘hand’

locative /kaː, H, L/ /uro, H, L/ /gɔː, LH, L/ /numo, LH, L/

kâː úrò gɔˇ ː` nùmôː

‘in the mouth’ ‘in the house’ ‘in the granary’ ‘in the hand’

Under the traditional assumption that all phonological features are binary specified properties of segments (Chomsky and Halle 1968), this is not additive but transformational exponence: The locative seems to transform a vowel with a feature specification encoding a H tone (say [+High –Low]) into the corresponding vowel which is specified for a falling tone (say [+High +Low]) (39a). Thus the data in (39) provide prima facie counterevidence to the Concatenativist Hypothesis. The revolutionary move of

inflectional exponence

69

Autosegmental Phonology and morphology initiated by Goldsmith’s dissertation on tone (Goldsmith 1976) was the assumption that features are entities which are in principle independent of segments (‘autosegments’) and might be located on separate layers (‘tiers’) of representation as shown in (40) for ka and go : (40)

Autosegmental representation of tone

This type of representation allows one to capture the locative as purely additive, namely as the addition of a H feature to the right edge of the tonal tier of its base under the assumption that tones which are not underlyingly associated to segmental material are associated with appropriate material by a general phonological process: (41)

Tonal locative in Jamsay as Additive Exponence

Jamsay illustrates two important points: First whether a given exponence pattern is additive, transformational, or templatic is not an inherent property of the data but is highly dependent on the phonological representations chosen by the analyst and the operations. Thus the Concatenativist Hypothesis might be more succinctly paraphrased as the claim that all (apparently) non-additive exponence is due to the choice of the wrong representations and rules, and will turn out to be additive under the appropriate choice of formalism. Second, this research programme stands or falls with the added value provided by choosing a phonological formalism which allows the additive reanalysis of apparently non-additive patterns. In the case of tone and the Jamsay locative, this added value is obvious. The autosegmental representation of tone is independently motivated by an ample amount of purely phonological phenomena such as the stability of tone under deletion of segments. Moreover, the autosegmental/additive approach considerably simplifies the analysis of the Jamsay locative itself: Thus if tone is specified as binary features of segments, it is difficult to model the fact that the locative transforms noun-final H vowels into falling ones, and R ones into a rise–fall, and that the H tone of the noun gets L if it is preceded by another syllable which is also H. Thus the segmental/transformational approach to tone would probably require three different rules of exponence. In the autosegmental approach Jamsay can be derived by a single exponence operation which is suffixal and phonological principles/rules which ensure that the floating tone is associated, the stem tone is maintained and (if possible under these premises) contours are avoided.

70

jochen trommer and eva zimmermann

Whereas the additive reading for tonal morphology of the type instantiated by the Jamsay locative in terms of autosegmental representations is by now a mainstream position, it is much more controversial whether all types of potentially non-additive exponence can (and should) be reinterpreted in this way. Bermúdez-Otero (2012) calls the hypothesis that non-additive exponence is generally triggered by the affixation of material which is in some sense non-segmental: Generalized Nonlinear Affixation (GNA). Non-segmental subsumes suprasegmental, subsegmental, or defective segmental (e.g. featurally underspecified consonants) phonological elements. Generalized Nonlinear Affixation forms a natural unit with the Concatenativist Hypothesis, and here, we will consider both assumptions as two sides of the same coin. In the following, we discuss how Generalized Nonlinear Affixation extends to templatic and to transformational exponence. Generalized Nonlinear Affixation and Templatic Exponence: Autosegmental Phonology also provides natural means to capture templatic exponence. Thus the Arabic binyan II (‘causative’) which is expressed by the templatic shape of a single consonant + vowel + geminate consonant + vowel + single consonant (cf. kattab, ‘cause to write’ based on katab, ‘write’) is captured in McCarthy (1979) by an affix consisting of a sequence of abstract timing units, C (consonant) and V (vowel) elements (CVCCVC) which are linked after affixation by general phonological rules to the vowels and consonants of the base verb. The same general approach is extended to reduplication with templatic elements by Marantz (1982). Thus the progressive reduplication in Mokilese (12), repeated in (42), produces not a complete copy of the base, but copies just enough material from the base to produce a heavy syllable, which amounts usually to a consonant–vowel–consonant sequence: (42)

Reduplication in Mokilese pɔdɔk mwiŋe kasɔ poki

progressive pɔd- pɔdɔk mwiŋ- mwiŋe kas- kasɔ pok- poki

‘plant’ ‘eat’ ‘throw’ ‘beat’

In the approach of Marantz, this can be captured by affixing a CVC prefix to a base (e.g. CVC-kC aV sC ɔV ) which is accompanied by copying the melodic material (the segments) of the base (CVC-kasɔ-kC aV sC ɔV ), association of the timing units to the segments of the copy (kC aV sC ɔ-kC aV sC ɔV ), and subsequent deletion of segments which are not linked to timing units (kC aV sC -kC aV sC ɔV ). Prosodic Morphology (PM) modifies the additive analysis for templates by assuming that templatic affixes which trigger copying and association are not CV or X-slot elements, but prosodic units such as syllables, feet, and moras. Under this account, the Mokilese progressive affix could be captured as a heavy syllable (σµµ ) which is then ‘filled’ by the segments of the copied base. Independently of the details of representation, templatic reduplication under this account

inflectional exponence

71

becomes additive exponence since a fixed amount of phonological material (a syllable or a CV-skeleton) is added to the base that results in a form that (partially) fulfils specific templatic requirements. In the next section, we will see evidence for the prosodic variant of the Mokilese analysis, and more generally for the take Generalized Nonlinear Affixation provides on Non-concatenative Allomorphy (NCA; cf. Section 4.3.1). GNA and Non-concatenative Allomorphy: In fact, in Mokilese, different reduplication patterns alternate to express the progressive. Thus for verbs without an initial CVC-sequence, the reduplicant is still a heavy syllable, but syllable weight is achieved by lengthening the vowel of the copied material: (43)

CVː-reduplication in Mokilese wia daul

progressive wiː-wia daː-dauli

‘do’ ‘pass by’

A transformational account would have to treat CVː-reduplication and CVCreduplication as two arbitrarily different allomorphs for progressive exponence whereas they can be unified to the single affixation of σµµ under the PM-analysis. Thus Mokilese also provides a case of Non-concatenative Allomorphy. An even more striking example for NCA which lends to a GNA interpretation is provided by Kirchner (2010) from the Wakashan language Kwak’wala, where the distributive is expressed by CV-reduplication if the base starts with a heavy syllable (44a), but by lengthening of the first base vowel if the first syllable is heavy (44b): (44)

Distributive reduplication in Kwak’wala a.

b.

stem wən maː c’aːs k’əmƛ c’əx k’əxw t’əs

distributive wənwəmuːt maːməmuːt c’aːc’əsmuːt k’əmk’əɬm’uːt c’aːxm’ʔuːt k’aːxwm’uːt t’aːsm’uːt

(Kirchner 2010: 15)

‘drill with auger’ ‘(serpent) crawl’ ‘eel grass’ ‘adze’ ‘singe’ ‘suck with whole mouth’ ‘crack barnacles’

As Kirchner argues, this allomorphy follows from affixation of a floating mora. The phonology prefers realization of this mora by associating it to the stem vowel but this is blocked if the syllable is heavy. This style of analysis also extends to the Saanich continuative allomorphy discussed above in (36), where either CV-metathesis, CV-reduplication, or ʔ-epenthesis mark the continuative aspect. This allomorphy is analysable as affixation of a mora since all three phonological operation add syllabic weight (Davis and Ueda 2006; Stonham 2007; Zimmermann 2009). Crucially, GNA is substantially more restrictive in its empirical consequences for Non-concatenative

72

jochen trommer and eva zimmermann

Allomorphy than Realize Morpheme. GNA predicts that NCA should always involve a fixed phonological target (such as in the Saanich case a mora), whereas for Kurisu’s Realize Morpheme, every phonological change should be in principle equivalent, hence vowel lengthening could in principle also alternate with completely unrelated phonological changes, say the deletion of a consonant. GNA and transformational exponence: A standard approach to segmental overwriting in Autosegmental Phonology is the addition of floating features to a base which leads to automatic phonological association to a base segment and concomitant deassociation of underlyingly associated features. Thus Lieber (1992) analyses voicing mutation in Aka as the prefixation of the feature [+voice]. See Akinlabi (1996); Zoll (1996); Wolf (2005, 2007) for reimplementations of this idea in OT, and Bye and Svenonius (2012) for a slightly different alternative using phonologically underspecified segments. However, there are different types of overwriting exponence which are more difficult for a GNA approach. First, Exchange Processes such as voicing polarity seem to be in principle incompatible with the Concatenativist Hypothesis since they do not involve a fixed phonological target (see Moreton 2004 for arguments that polarity is an impossible phonological pattern in standard OT). However, a survey by de Lacy (2012) shows that Dholuo is the only segmental exchange process discussed in the literature which provides enough evidence to be argued to be productive. But for Dholuo, different authors (Baerman 2007a; Kim and Pulleyblank 2009; de Lacy 2012) have argued recently that voicing exchange in Dholuo is at least partially the effect of phonological final devoicing, only roots with a final obstruent show apparent plural voicing (bat ⇒ bed-e, ‘arm(s)’, Okoth-Okombo 1982: 30). The same pattern is only marginally attested for vowel-final roots (agɔːkɔ ⇒ ag ɔ´ g-ˆɛ, ‘chest(s)’, Tucker 1994: 491). Second, a similar problem as with exchange processes is posed by overwriting processes which effect a Chain-Shift. Thus, in the so-called eclipsis pattern of Irish, a number of morphological categories are expressed by shifting voiceless obstruents to corresponding voiced ones and voiced oral stops to homorganic nasal stops. For example, all nouns preceded by a second or third person plural possessive pronoun or one of the cardinal numbers 7–10, undergo eclipsis in (45). (45)

Irish eclipsis radical t jax ‘a house’ f jiə ‘deer’ doːrəs ‘door’

eclipsed ə djax ‘their house’ ə vjiə ‘their deer’ ə noːrəs ‘their bag’

(Gnanadesikan 1997: 97, Pullman 2004: 13) [–vc] Stop

⇒

[+vc] Stop

[–vc] Fricative

⇒

[+vc] Fricative

[+vc] Stop

⇒

Nasal

inflectional exponence (46)

Irish eclipsis

73

(Gnanadesikan 1997: 97, Pullman 2004: 13)

However, eclipsis in Irish is analysed in Trommer (2011a) as ‘sonority affixation’, that is, the morpheme is taken to be a floating grid mark in multivalued featural representations of phonological scales. This makes the interesting prediction that chain-shifting segment mutation always leads to more sonorous (not to less sonorous) segments, a prediction confirmed by patterns of vowel quality chain-shifts in Western Nilotic (Trommer 2011a). Apparently the most obvious empirical problem for the Concatenativist Hypothesis are cases of subtractive morphology as in Tohono. However, Trommer and Zimmermann (2014) propose an analysis in terms of affixing a morphemic mora (= µ). The gist of the proposal is that the phonology of a language such as Tohono leads to association of the affix-µ to a base segment, but not to the higher prosodic structure (syllables, feet, etc.) of the phonological representation. To avoid domination of the base segment by multiple root nodes, it is deassociated from the dominating syllable and thus effectively deleted. This follows in an analysis based on the containment assumption (Prince and Smolensky 1993) which states that no underlying phonological element may be literally deleted. In such a framework, phonetic realization is bound to integration into higher prosodic structure (Stray Erasure in the sense of pre-OT phonology Steriade 1982, Itô 1988 and early versions of OT, Prince and Smolensky 1993): (47)

Subtractive morphology by µ-affixation

This analysis also captures the observation by Alber and Arndt-Lappe (2012) that subtractive morphology is generally templatic in the sense that the part of a base which is deleted corresponds to an invariable prosodic shape (in this case a mora). Impossible patterns under the Concatenativist Hypothesis: Apart from segmental exchange processes, the Concatenativist Hypothesis also predicts that rearrangement of base material should be an impossible pattern of exponence, a prediction which largely seems to be borne out. Although metathesis was discussed as one attested type of transformational exponence in Section 4.2.2, it is a striking observation that metathesis is

74

jochen trommer and eva zimmermann

never the only exponent of a morpheme. Virtually all cases where metathesis seems to be a productive means to express some morphological category are additive in the sense that metathesis is a side effect of adding weight to the underlying base, as we have already seen for Saanich. This is also true for metathesis in Rotuman where deletion of the final vowel, metathesis, umlaut, diphthongization, or no exponence alternate predictably to express the incomplete phase, one of two morphological ‘phases’ that are distinguished for every major class word in Rotuman indicating indefiniteness and non-emphasis (Churchward 1940; McCarthy 2000). Reordering of segments alone is consequently no exponence but only a side effect reanalysable as morphologically triggered prosodic weight adjustment. Apart from segmental metathesis, there are many conceivable types of reordering which seem to be unattested. Thus whereas there is limited evidence for phonologically triggered metathesis of tone (Hyman and Leben 2000), no language seems to employ the linear reordering of tone for purely morphological purposes. To take another conceivable option, imagine a language with a lexically determined stress system where the past tense of verbs is produced as shown in (48). If lexical stress is on the second syllable it shifts to the first one, if it is on the third syllable it shifts to the second, and if it is on the fourth to the third one. Thus stress moves one syllable to the left if possible: (48)

Unattested stress-shifting exponence present past

kéfatipa kéfatipa

patúlila pátulila

lirokámi lirókami

tokularú tokuláru

As far as we know there are no natural languages like (48). The best approximation are languages where exponence is achieved by stress shift which is templatic. Thus modern Greek forms its past tense by consistent antepenultimate stress (Holton et al. 1997): (49)

Fixed stress pattern in Greek past tense formation iðrío ‘I establish’ (disyllabic root √iðri-)

1sg 2sg 3sg 1plg 2pl 3pl

past active íðri-a íðri-es íðri-e iðrí-ame iðrí-ate íðri-an ∼ iðrí-ane

perfective past active íðris-a íðris-es íðris-e iðri´s-ame iðrís-ate íðris-an ∼ iðrís-ane

past passive iðri-ómun(a) iðri-ósun(a) iðri-ótan(e) iðri-ómastan iðri-ósastan iðrí-ondan ∼ iðri-óndusan

perfective past passive iðríθik-a iðríθik-es iðríθik-e iðriθík-ame iðriθík-ate iðríθik-an ∼ iðríθik-ane

Again, this pattern alternates with an additive pattern: If the root is monosyllabic the ‘augment e ’ is inserted:

inflectional exponence (50)

75

Epenthesis in Greek past tense formation √ ɣráfo ‘I write’ (monosyllabic root ɣráf-)

1sg 2sg 3sg 1plg 2pl 3pl

past active é-ɣraf-a é-ɣraf-es é-ɣraf-e ɣráf-ame ɣráf-ate é-ɣraf-an ∼ ɣráf-ane

perfective past active é-ɣraps-a é-ɣraps-es é-ɣraps-e ɣráps-ame ɣráps-ate é-ɣraps-an ∼ ɣráps-ane

past passive ɣraf-ómun(a) ɣraf-ósun(a) ɣraf-ótan(e) ɣráf-ómastan ɣráf-ósastan ɣráf-ondan ∼ ɣraf-óndusan

perfective past passive ɣráftik-a ɣráftik-es ɣráftik-e ɣraftík-ame ɣraftík-ate ɣráftik-an ∼ ɣraftík-ane

This is captured by van Oostendorp (2012) by affixation of an empty foot which seeks to remain as far as possible to the left edge and whose head is filled by e if this does not lead to a stress pattern which is illicit in Greek (stress more leftwards than the antepenultimate syllable), and is otherwise filled by the first two syllables of the base which results automatically in stressing the first syllable and suppressing the underlying lexical stress of the base7 since prosodic words in Greek contain maximally a single foot. Thus apparent stress shift is here actually overwriting of the underlying stress/foot structure by the affix foot.

4.4 Linear Order

.............................................................................................................................................................................

In the simplest of all worlds, linearization of inflectional material would work just as linearization of syntactic constituents by juxtaposition to the phonological material of the base. In fact the classic survey of Greenberg (1963) (cf. also Ultan 1975) finds that both prefixes and suffixes are much more frequent in the languages of the world than all other possibilities of linearizing inflection (especially infixation and interdigitation, cf. subsection 4.4.4). Slightly generalizing this result, one may formulate the hypothesis 7 Determining the underlying lexical stress of Greek verbs is a non-trivial task: Verbs never occur as bare roots (without affixal inflection), and inflection in turn has a crucial impact on the positioning of stress, not only in past tense forms, but also in the present where stress options are restricted to the final syllable of the root in the first conjugation (cf. aláz-u-me, ‘we change’) and the theme vowel in the second conjugation (cf. aɣap-á-me, ‘we love’, van Oostendorp 2012: 11). Thus it might well be the case that specific instances of underlying stress are also neutralized by the morphophonological context in present tense forms (e.g. in polysyllabic verbs where the underlying stress is on a non-final syllable). Still membership in the first and second conjugation is lexically determined, and leads to different (phonologically unpredictable) positions of stress in present tense forms, whereas past tense forms do not exhibit a lexical difference between verbs (or conjugation classes). Thus it is safe to conclude that at least some verbs have underlying specifications for the position of stress and that this position is overwritten in past tense forms.

76

jochen trommer and eva zimmermann

in (51), which underlies (albeit usually implicitly and in modified forms) much work on the morphology–phonology interface; (51)

The Strict Segmental Adfixation Hypothesis: All affixation is segmental adfixation to the segments of the morphological base (where Adfixation = {Prefixation, Suffixation})

The central empirical problem for (51) is the great variety of infixal exponence for inflection in the languages of the world, which will be the main focus of this section. In the following two subsections, we discuss the two prominent attempts to reconciling the Adfixation Hypothesis with the existence of infixational exponence, first maintaining it and deriving infixation by phonological mechanisms (subsection 4.4.1), and second, relaxing it in various ways (subsection 4.4.2). In subsection 4.4.3, we extend our survey to non-additive infixation and linearization, and subsection 4.4.4 introduces further patterns of exponence affecting linear order. Subsection 4.4.5 discusses remaining problems for the hypothesis in (51).

4.4.1 Deriving Infixation Phonologically The analysis which most straightforwardly captures the idea of implementing infixation as a strictly phonological process is the one of Horwood (2002) for the Tagalog actor focus exponent um.8 Crucially, um is not strictly an infix: it appears as a prefix with V-initial bases, but after the first C with consonant-initial ones: (52)

Tagalog um-infixation base abot tawag

actor focus umabot tumaawag

‘reach for, pfv’ ‘call, pfv’

In Horwood’s analysis, um is morphologically a prefix which appears at the left edge of its base just as any other prefix. V-initial bases: that um does not ‘move’ into its base in the phonological output is ensured by the constraint Linearity which penalizes reverting the order of segments (53). However the constraint NoCoda which codes the cross-linguistic dispreference for syllables with coda consonants is ranked above Linearity. With C-initial bases, strict prefixation of um leads to an additional violation of NoCoda (by the final C of um ) (54b), which may be avoided by positioning 8

Horwood’s analysis is built on intuitions of the better-known analysis of Tagalog infixation in McCarthy and Prince (1993a), where infixation is not strictly metathesis, but violation of morphemespecific constraints requiring leftmost or rightmost position for affixes. We discuss Horwood’s analysis here because it relies much less on OT-specific technology.

inflectional exponence

77

um after the first C of the base (54a). Still, linear rearrangement of um is minimal: it remains as close to its original position as possible. (53)

Infixation in Tagalog: V-initial base um-abot ☞

(54)

a. u.ma.bot b. a.um.bot c. a.bu.mot d. a.bo.umt e. a.bo.tum

NoCoda

Linearity

* *!* * * *

* *!* *!** *!***

Infixation in Tagalog: C-initial base um-tawag ☞

a. um.ta.wag b. tu.ma.wag c. ta.um.wag d. ta.wu.mag e. ta.wa.umg f. ta.wa.gum

NoCoda

Linearity

**! * **! * * *

* ** **!* **!** **!***

Thus in Horwood’s analysis, infixation is literally phonologically triggered metathesis (cf. also Halle 2001). There are a number of drawbacks with Horwood’s analysis. First, infixation cannot always be predicted from the phonological shape of affixes and their bases. Thus Ilokano, another Austronesian language closely related to Tagalog has a similar pattern of um-infixation (55a), but also a prefix ag which is phonotactically isomorphous to um, and does not infix (55b) (Vanoverbergh 1955: 132, 137; Zoll 1996; Yu 2002): (55)

um vs. ag in Ilokano a.

b.

base isem kagat

um-form umisem kumagat

‘(threatens to) smile’ ‘(threatens to) bite’

base isem kagat

ag-form ag-isem ag-kagat

‘(actually) smiles’ ‘(actually) bites’

Thus Horwood would have to assume that there are actually morpheme-specific (indexed) versions of Linearity, where Linearityum is ranked below, and Linearityag above NoCoda (see Section 4.3.1 for critical discussion of indexed

78

jochen trommer and eva zimmermann

constraints). Another complication is that metathesis in Tagalog and Ilokano is obviously restricted to heteromorphic segments. Thus Tagalog *mutawag, where the segments of um itself are rearranged would otherwise be preferred since it only changes the linear order of u and m, whereas tumawag rearranges u with respect to k, and m with respect to k. But probably the most grave problem with this style of analysis is that it inherently predicts that infixation always optimizes phonological representations. As argued in extensive cross-linguistic detail by Yu (2007), there are a number of striking counterexamples to this claim. A typical case is the nominalizing infix ni in Leti, an Austronesian language spoken on the island of Leti. The affix appears consistently after the first C of a C-initial base even though this makes syllable structure worse, not better. Thus the putative form *ni-kaːti (56a) avoids the complex onset of k-ni-aːti, and has otherwise the same amount of codas and onsets. Since complex onsets are well-established cases of phonologically marked structure, the Horwood approach predicts that the Leti marker should not infix. (56)

Nominalizing -ni- in Leti a. kaːti b. kasi c. kakri

‘to carve’ ‘to dig’ ‘to cry’

k-ni-aːti k-ni-asi k-ni-akri

(Blevins 1999) ≪ ≪ ≪

*ni-kaːti *ni-kasi *ni-kakri

‘carving’ ‘act of digging’ ‘act of crying’

4.4.2 Relaxing the Adfixation Hypothesis—Anchor Points and Pivots The alternative approach to infixation proposed by Yu (2007) is based on the assumptions that prefixation and suffixation may morphologically arbitrarily target a specific member of a set of cross-linguistically possible anchor points (‘pivots’) given in (57) (see Fitzpatrick 2004 for a slightly different inventory of anchor points): (57)

Possible pivots for affixation

(Yu 2007)

a.

Initial pivot (i) First consonant/onset (ii) First vowel/nucleus . (iii) First syllable b. Final pivot (i) Final vowel/nucleus (ii) Final syllable c. Prominence pivot (i) Stressed syllable (ii) Stressed vowel/nucleus Note that this approach still embraces part of the assumptions of the Adfixation Hypothesis: Affixes are either prefixes or suffixes, not to the entire base but to a specific pivot of the base.

inflectional exponence

79

Thus Tagalog um is a prefix to the first vowel of its base, whereas the Alabama mediopassive marker l in (58) is a suffix to the same pivot (there is a regularly epenthetic i before consonant clusters in (58a, b)): (58)

Alabama mediopassive: a suffix to the leftmost vowel (Martin 1994; Yu 2007) a. tak.co b. hoc.ca c. oti d. a-hica

‘rope(verb)’ ‘shoot’ ‘make a fire’ ‘watch over’

talikco holicca oːlti a-lhica

‘be roped’ ‘be shot’ ‘kindling’ ‘be taken care of ’

Thus Yu still adheres to the hypothesis in (59): (59)

General Adfixation Hypothesis: All affixation is adfixation to some phonological constituent

Yu’s theory makes a further prediction which is in contradiction to the Horwood-style phonological approach to infixation: Infixation is expected to be fixed, that is, an infix cannot appear at different anchor points in bases with different phonological shape. In contrast, it is conceivable under the phonological approach that an infix appears after the first consonant of some bases, but after the second consonant of other bases, if this distribution maximizes phonological well-formedness. Note finally that ‘infixes’ in Yu’s theory are also not necessarily consistent infixes. Thus if an affix targets the first vowel of the base it might be still a prefix if the base is V-initial and an infix if the latter is C-initial.9

4.4.3 Linearization of Non-additive Exponence Reduplicative infixes seem to target roughly the same locations as purely additive infixes. Thus Quileute, a Chimakuan language of the Olympic Peninsula in Washington, has a reduplicative CV-infix expressing plural which always occurs after the first vowel of the base similarly to the Alabama example discussed above (58): (60)

Reduplicative infixation in Quileute (Andrade 1933: 188; Broselow and McCarthy 1983: 44) a. b. c.

sg qaːwats t’aːdax k’aʔt’la

pl qaːqeːwats t’aːt’eːdax k’ak’eːʔt’la

‘potato’ ‘tail (of a bird)’ ‘stone’

9 See Yu (2007) for the discussion of ‘true’ infixes (affixes which occur infixed with all bases); Yu explains such infixes by double pivots: the affix selects a pivot on its left and also one to its right.

80

jochen trommer and eva zimmermann

An interesting asymmetry in the distribution of reduplicative exponence is that there is a strong typological preference for reduplication to appear in prefixal position (cf. Marantz 1982) which runs against the global preference for suffixation (Greenberg 1963; Cutler et al. 1985; Bybee et al. 1990).10 Nelson (2003) explains this by the interaction of a locality constraint corresponding to ‘Marantz’s generalization’—in the unmarked case, reduplicating prefixes copy leftmost substrings of their bases (e.g. badupi ⇒ ba-badupi ), and reduplicating suffixes rightmost substrings (e.g. badupi ⇒ badupi-pi ) (Marantz 1982: 447) and the requirement that the left edge of reduplicative affixes is identical to the left edge of the base. A suffixal reduplicant which copies only parts of the base (e.g. because it is restricted to syllable size) can never satisfy both requirements at the same time for a longer base; it either fails to copy the left edge of the base (e.g. badupi ⇒ badupi-pi ) or copies the left edge non-locally (e.g. badupi ⇒ badupi-ba ). Also other types of transformational exponence appear in places typical of segmental affixes or infixes: Consonant mutation typically targets initial or final segments of its base. Apophony often targets leftmost or rightmost vowels, but is also often bound to stressed vowels. Similarly morphological vowel lengthening often affects the rightmost vowel of its base (cf. Quechua as discussed in Kirchner 2007), or is infixal (cf. the Shizuoka Japanese data in Davis and Ueda 2006). The template in Cupeño might be argued to be suffixed to the stressed vowel of its base. However, in contrast to segmental infixes there are clear cases of infixal exponence affecting segmental features which exhibit different positions according to the phonological shape of the base. The most famous example for this is labialization in the Ethiosemitic language Chaha which is part of the inflectional marking of 3sg.m objects on verbs, and appears at the rightmost C in the base (if any) which can be labialized, that is, the rightmost noncoronal. Thus in (61a), the last C is labialized, in (61b) the penultimate C (the last C is coronal), in (61c) the antepenultimate C (the last two Cs are coronal), and in (61d) no C is labialised because all root Cs are coronal: (61)

Labialisation of dorsal and labial Cs in Chaha a. b. c. d.

verb dänäg näkäs qätär sädäd

verb + 3sg.sm objekt dänägw näkwäs qwätär sädäd

(Rose 2006; Akinlabi 1996)

‘hit’ ‘bite’ ‘kill’ ‘chase”

On the other hand, there are cases where an expected infixation of non-segmental exponence is blocked in some contexts. In Bukusu, a Bantu language, two different types of exponence co-occur in the formation of the first person singular. As can be seen in (62a), the subject prefix n- is accompanied with lengthening of the vowel in the following syllable. Under the Concatenativist Hypothesis, this can be analysed as an 10 As noted by Hyman (2008: 316) a similar prefixing preference (left-edge orientation) also holds for infixation.

inflectional exponence

81

affix that consists of a segmental representation + a floating mora. This floating mora must attach to some segment in order to be properly realized. (62)

Lengthening in Bukusu a.

b.

stem ixala esa asama ora lima taβula kona cˇ uxa

1sg niːxala neːsa naːsáma noːra ndíma ndáβula ngóna njúxa

(Mutonyi 2000) ‘sit’ ‘hangout’ ‘open mouth’ ‘bask’ ‘cultivate’ ‘tear up’ ‘sleep’ ‘pour/spill’

But this analysis is faced with serious problems as soon as one takes a look at forms with an initial onset as in (62b). In such contexts, no lengthening can be observed. The onset consonant that intervenes between the prefix and the base vowel apparently blocks the lengthening. This blocking effect is intuitively plausible but cannot easily be integrated in standard moraic theory: onset consonants are standardly taken to be non-moraic. Consequently, nothing intervenes on the moraic tier between the lengthening mora of the prefix and the stem vowel. The adfixation hypothesis is therefore incapable of explaining the absence of lengthening in these cases. It rather seems that phonological elements on quite different tiers are able to block affixation of elements on another tier. (63)

Affixal lengthening in Bukusu a. Affixal lengthening

b. No affixal lengthening

Mutonyi (2000) argues that this phenomenon in Bukusu can only be captured in a theory where skeletal elements are arrayed on the same tier as the onset: hence for X-Slot theory. A final asymmetry between segmental and non-segmental exponence is the observation made by Hyman (2001) that there seem to be no tonal infixes.

82

jochen trommer and eva zimmermann

4.4.4 Further Types of Exponence Affecting Linear Order As we have already mentioned, metathesis as a means of inflectional exponence is only attested marginally, it is never unconditional (it occurs only with bases of specific phonological shapes), always stands in allomorphic relation with other means of exponence, and has additive effects (by introducing additional moras). Circumfixes may be considered the converse of infixes, but the mere existence of circumfixes is subject to debate, since the phenomenon is relatively rare, and in most alleged cases of circumfixes at least one of the parts (the prefix or the suffix part) also occurs independently of the other one (cf. Spencer 1991: 13; Klamer 2005). Moreover, any theory which allows for multiple exponence for the same inflectional features (e.g. by two suffixes in different position classes) implicitly allows one to model circumfixes as combinations of a prefix and a suffix. Thus assuming circumfixation as an additional logically independent mode of inflectional exponence seems to induce systematic analytic ambiguity. Interdigitation, as observed in Semitic languages has also been sometimes referred to as an independent way of linearizing inflectional exponence (cf. the perfect passive of Classical Arabic which is formed by replacing the vowels of a verb stem by u and i, e.g. katab ⇒ kutib ), but this has been argued to involve a case of multiple infixation (Ussishkin 2003, 2005).11 Note finally an interesting typological observation by Marušiˇc (2003). Marušiˇc argues that there are no attested cases of disjoint affixes, that is, an affix which appears as a continuous string of segments with some bases (of specific phonological shape), but as a combination of an adfix and an infix, that is, discontinuously with other bases, as shown for the hypothetical affix ka in (64): (64)

a. napa-ka

b. ma-k-tond-a

4.4.5 Other Problems for the Adfixation Hypothesis The Adfixation Hypothesis in all the forms discussed in this chapter comprises the assumption in (65): (65)

The Strict Ordering Hypothesis: Affixes target morphologically a single fixed position (left or right) with respect to their phonological bases

The most striking counterexample to (65) are mobile affixes, which occur as prefixes with some bases, but as suffixes with others. Thus in Huave (Kim 2008: 332), the stative affix n and the completive affix t occur as prefixes with vowel-initial, but as suffixes with C-initial stems:

11 Trommer (2005, 2008b) shows that the vowel patterns in Amharic, another Semitic language actually consist of different inflectional markers. Thus every vowel is an independent infix.

inflectional exponence (66)

83

Mobile affixes in Huave (Kim 2008: 332) a.

b.

v-initial stem n-[a-kants] st-tv-red ‘red’ t-[u-ty] cpl-tv-eat ‘(s)he ate’

c-initial stem [pal-a-n] close-v-st ‘closed’ [mojk-o-t] face.down-v-cpl ‘(s)he lay face down’

As argued by Kim, this distribution is at least partially phonologically optimizing. Thus realizing n or t as prefixes with C-initial stems would lead to complex onset clusters (e.g. *n-pal-a ).

4.5 Conclusion

.............................................................................................................................................................................

In our eyes, work on inflectional exponence is at a critical point. On the one hand, the research of the last decades has provided a thorough understanding of possible exponence patterns (see Section 4.2), and developed well-articulated theoretical research programmes which make clear-cut empirical predictions (see Sections 4.3 and 4.4). But this success in turn highlights the fact that for many of the more complex and theoretically decisive exponence types such as metathesis and polarity, it is not obvious whether grammatical descriptions of these phenomena in the literature provide theoretically significant data or refer spurious generalizations on morphophonological phenomena of a different type (cf. our discussion of Saanich and Dholuo). A related problem is that it is often not clear if a descriptive pattern of inflectional exponence is productively applied by speakers of a language or just reflects a collection of memorized suppletive forms inherited by historical accident. A case in point is umlaut and ablaut in Germanic (cf. Section 4.2.1) for which there is an ongoing controversy on their status as a morphological rule (see e.g. Prasada and Pinker 1993; Albright and Hayes 2003; Bermúdez-Otero 2012, on strong verbs in English). Thus it is a major challenge for future research to develop both the range and the methods for empirical research in this area.

p a r t ii ........................................................................................................

PARADIGMS AND THEIR VARIANTS ........................................................................................................

chapter 5 ........................................................................................................

INFLECTIONAL PARADIGMS ........................................................................................................

james p. blevins

5.1 Introduction

.............................................................................................................................................................................

The status of paradigms varies widely across approaches, ranging from the PostBloomfieldian view that they are ‘epiphenomena’, to classical word and paradigm (WP) conceptions of paradigms as fundamental units of lexical organization. This variation rests in turn on different views of the structure of morphological systems. Within the Post-Bloomfieldian tradition, paradigms represent collections of forms which are based on a common root and/or which have a partially overlapping derivational history. In line with the rigidly syntagmatic perspective of these accounts, it is the relation between formatives and words (or larger units) that determines the structure of a morphological system. The sets defined by the inflected (or derivationally related) forms of an item have no grammatical status or coherence as ‘units’. At the other extreme are classical WP approaches, which factor grammatical systems into sets of exemplary paradigms and inventories of principal parts. Between these extremes lies a range of approaches that incorporate a notion of ‘paradigm’ that corresponds to sets of forms, or to sets of abstract ‘cells’, which are usually defined in terms of feature bundles. This chapter sets out some of the dominant conceptions of paradigms that have been developed in different morphological traditions. Section 5.2 identifies the kinds of elements that have been organized into paradigms, and contrasts alternative notions of paradigmatic structure. Section 5.3 summarizes types of relations that have been argued to require reference to a notion of paradigm. Section 5.4 reviews recent applications of information theory to model the implicational structure of paradigms. Section 5.5 concludes by considering how conceptions of paradigms reflect general assumptions about inflectional systems.

88

james p. blevins

5.2 The Organization of Paradigms

.............................................................................................................................................................................

In parallel to ontological questions about the types of units that make up paradigms, there are issues concerning which units of whatever kind are grouped together into paradigms and about the organizational principles that bind these units into a structure. Inflectional paradigms are the core case of morphological paradigms, though some accounts develop a notion of derivational paradigm or ‘morphological family’ (de Jong 2002) that groups derivationally-related forms or items. Within the class of inflectional paradigms, there are different sets of cells or forms that can be identified as constituting sub-paradigms. There is relatively little ambiguity in most declensional systems, since the cells or forms of a nominal item can usually be organized into a single matrix of elements that differ for the features (case, number, gender, definiteness, etc.) that are distinctive in a given language. Yet even nominal paradigms may exhibit splits of various kinds. The paradigm of an item can be organized into sub-paradigms, whose members are more closely associated with each other than with members of other sub-paradigms. German provides a simple case of this type of split, as discussed in Section 5.3.2. Singulars and plurals each form sub-paradigms whose members are more similar to each other than they are to members of the other sub-paradigm. In languages with a dual, such as Slovene, the dual and plural forms likewise tend to show a stronger affinity to each other than to singulars. In declensional systems with large case inventories, such as Finnish or Estonian, it is common for there to be a split between the core ‘grammatical’ cases and the set of ‘local’ or ‘semantic’ cases. In a language such as Tundra Nenets, nominal paradigms may be organized into multiple clusters, each based on a common stem (Ackerman et al. 2009). The even larger form inventories found in Daghestanian languages are likewise organized into multiple ‘series’ of local cases (Kibrik 1998). However, there is some debate about the precise number of case distinctions in these languages (Comrie and Polinsky 1998), and some uncertainty about the combinatorics of spatial series markers and case endings. In each of these cases, the traditional paradigm represents a unit of organization between the forms in a sub-paradigm and the full morphological family of an item. Conjugational systems add a further level, since the forms that define a verbal lexeme may be organized into multiple paradigms, defined by different tense, aspect mood, and voice features.1 The fact that forms tend to be more strongly associated within than across sub-paradigms underlies the classical WP practice of describing paradigms in terms of multiple principal parts. 1 There is a fairly robust consensus for organizing conjugational paradigms into subsets with common tense/aspect/mood/voice features and variable agreement features, as opposed to an inverse organization into sets of forms that share common agreement features and have contrasting tense/aspect/mood/voice features. Hence it is more customary to talk of ‘the paradigm’ of a verb or ‘the present indicative (sub-)paradigm’ than of ‘the first person singular paradigm’.

inflectional paradigms

89

In sum, a form in a given language may be assigned to successively larger collections, from sub-paradigms through to morphological families. The familiarity of the inflectional paradigm, in the sense of the complete set of cells or forms of an item, is due in large part to the relevance of this notion to grammatical descriptions and the usefulness of the notion for pedagogical purposes. The structure associated with paradigms tends to depend fairly directly on assumptions about the types of elements that are grouped together into paradigms. Approaches that treat paradigms as sets of forms often define this set in terms of derivation from common bases. The model of A-Morphous Morphology (Anderson 1992) is typical in defining the paradigm of an item as the ‘set of surface word forms that can be projected from the members of its stem set’ (1992: 134), where the stem set contains the root of an item and any stem allomorphs. Approaches that treat cells as abstract feature bundles regard paradigms in turn as n-dimensional matrices, in which each dimension is defined by a distinctive feature. Within models such as Paradigm Function Morphology (PFM) (Stump 2001b), this abstract paradigm space is defined by the feature distribution constraints of a language, independent of the forms that come to be associated with cells by ‘spell-out’ rules.2 The fact that A-Morphous Morphology and Paradigm Function Morphology are both realizational models shows that the precise characterization of paradigms is something of a free choice in this class of models. Assumptions about the organization of paradigms are more deeply embedded in the analytic practices of classical WP models. This tradition proceeds from ‘the general insight that one inflection tends to predict another’ (Matthews 1991: 197), with the primary locus of prediction identified as ‘words as wholes, arranged according to grammatical categories . . . distinguished by their endings’ (1991: 187).3 Paradigms consist of cell-form pairs that are related not by shared bases or derivational histories but networks of interdependencies. In some cases, sets of interpredictable cells may be morphosyntactically coherent. However, they need not be, and in the patterns that Matthews (1972, 1991) terms ‘parasitic’ and Aronoff (1994) calls ‘morphomic’, the features of interdependent cells may serve solely to pick out pairs of forms that alternate or covary in some way. The classic WP factorization of inflectional systems into principal parts and exemplary paradigms embodies a strong hypothesis about the interdependency of forms in a paradigm. A language whose paradigms consisted entirely of mutually independent forms could not be broken down in this way, since no set of forms smaller than a full paradigm would identify the class of an item. This holds independently of the complexity of a system, as one can see by considering a system with just two cells, each associated with two distinct patterns of exponence. If each pattern of exponence that 2 Later elaborations of PFM introduce a split between ‘form paradigms’ and ‘content paradigms’, in part to describe patterns of heteroclisis. See Stump (2006). 3 Although Matthews (1991) here specifies ‘endings’ in the context of a discussion of Latin, predictive relations are understood to apply to any systematic patterns of form variation.

90

james p. blevins

realizes the first cell can co-occur with both of the patterns that realize the second cell, the system will contain a total of four patterns. There is no difficulty in exemplifying paradigms, associating them with classes, or in assigning individual items to classes. However, this classification is purely taxonomic, and the paradigms and classes it defines cannot play the role that exemplary paradigms do in a classic WP analysis. Since the forms of an item are inflectionally independent, the system does not define principal parts, and each item must be represented by both forms. If a system has weakly interdependent forms, then a factorization into exemplary paradigms and principal parts may also achieve negligible economy and merely distribute patterns across a large and unilluminating collection of paradigms that ‘multiply out’ the possibilities of co-occurrence. Hence patterns of mutual implication in a system must be relatively tight in order for a classic WP factorization to simplify rather than complicate the description of a system. There is an intimate connection between the assumptions that an approach adopts about the elements in a paradigm and the organization imposed on those elements, and the usefulness of the resulting structure. Sets of simple forms are relatively uninformative; the members of such form sets can be organized in terms of their morphotactics or their derivational history, but shared morphotactic or derivational structure rarely play any direct grammatical role. So it is unsurprising that little if any grammatical significance is attached to paradigms in models that treat them merely as sets of simple forms. Sets of abstract cells intrinsically define a much richer structure. It is within this essentially closed and uniform space that notions like ‘suppletion’, ‘syncretism’, and ‘paradigm gap’ are defined most clearly. The features that define this space likewise provide the vocabulary for the markedness proposals of Jakobson (1932, 1936), and the property co-occurrence restrictions and other types of constraints developed in approaches such as Stump (2001b). However, pairings of forms and abstract cells define the most informative elements, and determine the richest structure, since relations may refer both to the features and to the form of a cell. Implicational relations highlight the contrast between these conceptions of paradigms. Words or sub-word units are of limited predictive value qua forms. Knowing that the paradigm of an item contains a particular word form is often relatively uninformative unless one knows what cell the form realizes and, consequently, one knows where the form fits into the paradigm. Abstract cells support predictions based on markedness or other types of relations between unrealized features. Pairings of forms and abstract cells are again of the greatest predictive value, as they sanction deductions between forms and between cells but also deductions between forms based on their place in a paradigm. In setting out the evidence for treating paradigms as central components of the inflectional system of a language, this chapter will attempt to present the strongest case for a classic WP perspective by adopting the most informative conception of paradigms, one in which they are structured sets of form–cell pairs. As in the case of words, the role of paradigms in pedagogical and reference grammars attests to their practical usefulness. The pedagogical relevance of paradigms reflects

inflectional paradigms

91

their descriptive value, at least for languages of the ‘flectional’ type (Sapir 1921), as Matthews (1991) emphasizes. Pupils begin by memorising paradigms. These are sets of words as wholes, arranged according to grammatical categories. This is not only traditional, it is also effective. They learn that different members of a paradigm are distinguished by their endings. . . . They can then transfer these endings to other lexemes, whose paradigms they have not memorised. . . . It seems unlikely that, if a structuralist method or a method derived from structuralism were employed instead, pupils learning Ancient Greek or Latin—or, for that matter, Russian, Modern Greek or Italian—would be served nearly so well. (Matthews 1991: 187f.)

Comprehensive descriptions of languages with intricate morphological systems also tend to exhibit exemplary paradigms and patterns at different levels of generality. No practical alternative has yet been devised that would replace exemplary patterns by something like morpheme inventories related to surface realizations by rules or constraints. In part, this reflects the fact that Post-Bloomfieldian models were not designed with practical or comprehensive description in mind, so that the output of their analyses may range from nontransparent to indeterminate (Karttunen 2006). This tradition is accordingly sceptical about the relevance of the pedagogical or descriptive utility of paradigms. It is of course true that the notion of paradigms represented in pedagogical and reference grammars is strongly influenced by the uses to which paradigms have been put in these sources. It is usually assumed that inflectional systems can be factored into a discrete number of inflection classes, each represented by the full paradigm of some exemplary member of the class. Principal part inventories are normally ‘static’ in the sense of Finkel and Stump (2007), in that each non-exemplary item of a given class is represented by the same forms or sets of forms, for example the nominative singular or first person singular. The deduction of new forms from exemplary paradigms and principal parts is likewise attributed to symbolic processes of the kind expressed by proportional analogies. Each of these assumptions creates problems that remain largely unaddressed, and mostly unacknowledged, in the descriptive and pedagogical sources that make use of paradigms. The difficulty of motivating the choice of principal parts (or ‘leading forms’) is perhaps the best known of these problems. One objection to the Priscianic model . . . was that the choice of leading form was inherently arbitrary: the theory creates a problem which it is then unable, or only partly able, to resolve. (Matthews 1972: 74)

The other assumptions raise similar problems. Although pedagogical traditions tend to agree broadly on the number of classes in a language (and even on the choice of exemplary members of each class), this consensus often reflects established practices

92

james p. blevins

or considerations of utility. There are simply no generally agreed criteria for class identification, let alone for selecting exemplary items. The optimal solution to these problems avoids them altogether by recognizing that they derive from practical idealizations rather than core properties of paradigms or paradigm-based models. For pedagogical purposes, it is useful to draw the most informative cells of a paradigm to the attention of language learners. However, there is no reason to assume that a single form will always identify the inflectional pattern of an item.4 Conversely, there is no reason to ignore the partial information supplied by other forms. There are also no grounds for assuming that a language can be organized into some fixed set of classes, independent of the purposes or level of detail of the classification. Instead, different sets of interdependent forms may be defined based on their predictive value. For pedagogical purposes, it will often be useful to bundle these sets of interdependent forms into larger collections that specify the shape of each form of a single item, irrespective of how loosely the larger forms of different sets are connected to each other. In most languages, it is of course likely that there is at least some member of each class whose forms are frequent enough that the full paradigm of that item could serve as an exemplar for the class. However, the contrast between forms that are taken to be resident in the mental lexicon of a speaker and those that must be deduced from previously-encountered forms will depend on frequency, and it is known that the different forms of an item may have very different frequencies of occurrence. There is also no principled reason to assume that the analogical processes that deduce new forms of an item should be representable symbolically, rather than sub-symbolically. What remains after these idealizations have been stripped away is a conception of paradigms as structured networks of interdependent elements. Although it is possible to generalize over these networks, traditional paradigm-based models incorporate a distinctively exemplar-based conception. The leading idea is that form variation within a morphological system is exhibited by exemplary patterns that serve a dual role within a grammatical system. On the one hand, the paradigms resident in a speaker’s mental lexicon (whether full or partial) specify forms of particular lexical items. On the other hand, these structures provide a model for the inflection of as yet unencountered forms of other items. The pedagogical practice of identifying a single exemplary paradigm for each inflection class represents an idealization over the sets of resident forms that collectively reinforce the patterns exhibited by the exemplary paradigms.

5.3 Paradigmatic Relations

.............................................................................................................................................................................

The value of this classical WP notion of paradigm depends on the types of generalizations and insights that it supports. Within the linguistic literature, a number of 4 Languages in which regular items can be represented by a single diagnostic form will of course conform to a pedagogical ideal of inflectional economy. A similar conception underlies more theoretical principles, such as the Single Base Hypothesis of Albright (2002).

inflectional paradigms

93

principles and constraints have been proposed that operate in a distinctive way within paradigms. Although not always directly formulated in terms of paradigms, these constraints exploit the fact that members of a paradigm tend to have numerous properties in common, including ‘inherent’ grammatical features, a core lexical meaning, and a shared stem, or stem set. These properties put the elements of a paradigm in close ‘competition’ for syntagmatic slots, which may give rise to grammatical reflexes. The most familiar reflex is illustrated by cases in which competition between alternatives is taken to be resolved in favour of the most specific alternative. This interaction is expressed by a range of principles, including Morphological Blocking (Aronoff 1976), the Elsewhere Condition (Kiparsky 1982b), P¯an˜ ini’s Principle (Janda and Joseph 1992; Stump 2001b), and the Disjunctive Ordering Principle (Anderson 1986). Another reflex of competition between forms of an item is a type of ‘paradigmatic deduction’ that allows a form to be interpreted based on the absence of a marker. Noun paradigms in English provide a simple example. Regular plural nouns in English are marked by the suffix /z/ (represented orthographically as -s). There is, however, no marking of singular number, and none is needed, given that a singular noun is unambiguously identified by the lack of a plural marker. A zero morph adds no information to what speakers can already deduce from the absence of any realized exponent.5 Similar patterns are even more typical of more intricate paradigms. As Anderson (1992) notes, Georgian verbal paradigms provide a striking illustration of the fact that ‘information may sometimes be conveyed not by constituents that are present in the structure of a given word, but precisely by the fact that certain other material is absent’. Consider the Georgian Verb form [dagxat’e] in [Table 5.1], for example, . . . This form represents agreement with a first-person singular Subject and a second-person singular Direct Object. . . . But while an overt affix (/g/) is present to signal agreement with the second-person Object, no affix marks the fact that the Subject of this Verb is (and must necessarily be) first-person singular. This agreement can be inferred from the following information. The Subject cannot be second person, because if it were, the sentence would be reflexive—but reflexive forms in Georgian are grammatically third person, and this Verb has a second-person Object. Similarly, the Subject cannot be third person, since, if it were, there would be a suffix [/-a/] at the end of the Verb. Thus . . . the Subject must be first person. But it must be singular, rather than plural, since a first-person plural Subject would trigger the introduction of a suffix /-t/) at the end of the Verb. We know therefore that the Subject of this Verb must be first-person singular, but this fact is not signaled by the presence of any overt affix in the word. (Anderson 1992: 87) 5

Early Post-Bloomfieldian accounts introduced a ‘zero’ singular marker in such cases (Harris 1942: 110), and the practice survives in some contemporary work. Hockett (1947: 230) recognized that this type of element had ‘a very dubious status’, and the coherence of a ‘zero marker’ is questioned by Matthews (1991: 124). From a classic WP perspective, the appeal to ‘zeros’ can be seen as a means of compensating for the rigidly syntagmatic character of the post-Bloomfieldian model, which has no provision for any type of paradigmatic comparison.

94

james p. blevins Table 5.1 Aorist indicative paradigm of XAT ’VA ‘paint’ Subject

1sg

1pl

Object 2sg

2pl

3

1sg 1pl 2sg 2pl 3sg 3pl

– – damxat’e damxat’et damxat’a damxat’es

– – dagvxat’e dagvxat’et dagvxat’a dagvxat’es

dagxat’e dagxat’et – – dagxat’a dagxat’es

dagxat’et dagxat’et – – dagxat’at dagxat’es

davxat’e davxat’et daxat’e daxat’et daxat’a daxat’es

Source: Tschenk´eli (1958: §31).

This traditional deduction interprets an inflected form in the context of its paradigm. The deduction is, as Anderson shows, effective, provided that two conditions are met. First, the speaker must be able to identify the paradigm of dagxat’e in order to compare this form against relevant alternatives. Second, given that the deductions involve whole forms, not localized patterns of exponence, the speaker requires access to the full aorist indicative paradigm of xat’va—or sufficient information to reconstruct a paradigm of this class.6

5.3.1 Paradigm Economy More direct evidence for paradigms takes the form of constraints or generalizations that apply to paradigms. In the diachronic domain, notions like ‘levelling’ and ‘extension’ are principally defined with reference to inflectional paradigms (Hock 1991). In the synchronic domain, various constraints have been proposed that exploit, directly or indirectly, the implicational structure that binds the cells of a paradigm into a cohesive whole. This structure underlies the organization and economy of classic WP models, as Matthews (1991) notes. The most general insight is one inflection tends to predict another. This insight can be incorporated into any model. Traditionally, it is the basis for the method of exemplary paradigms. If the alternations were independent, these would have to be numerous. (Matthews 1991: 197)

Matthews’ observation that predictive relations prevent alternations from becoming too ‘numerous’ also gives the classic WP explanation for what have come to 6 Interestingly, both conditions rely on a notion of paradigm that is a more coherent and cohesive structure than the simple set of output forms that Anderson (1992) proposes elsewhere.

inflectional paradigms

95

be known as ‘paradigm economy’ effects. The contemporary interest in paradigm economy effects stems largely from the discussion in Carstairs (1983), which draws attention to these effects and attributes them to a dedicated Paradigm Economy Principle (PEP). Subsequent elaborations of this approach attempt to subsume the effects of the PEP under a range of more general principles (Nyman 1986; Carstairs-McCarthy 1991, 1994). However the general conception of economy that underlies these proposals is articulated most clearly in the initial formulation of the PEP. In contrast to a traditional model, the PEP is not concerned with distinctive inflectional patterns in general but with ‘inflectional resources’, by which Carstairs (1983) means ‘affixal resources’. Paradigm economy provides at least a partial answer to a question which, so far as I can discover, has not been asked before—a question about how, in any inflected language, the inflexional resources available in some word-class or part of speech are distributed among members of that word class. (Carstairs 1983: 116)

The distribution of inflectional resources is characterized in terms of upper and lower bounds on the space of possible inflection classes, where these bounds are determined by the number of different affixal strategies available for realizing paradigm cells in a given word class. This can be illustrated by a schematic example. Consider a simple declensional system with two contrastive properties, number and case, two distinct number features, N1 and N2 and four case features, C1 , C2 , C3 and C4 . Since each combination of number and case features defines a paradigm cell, these number and case features define the family of eight-cell paradigms summarized in Table 5.2. The actual number of realizations for each cell are determined mainly by the size of the nominal lexicon, which is independent of the economy of the inflectional system. Hence to describe inflectional economy one must first isolate patterns of inflectional exponence. For each cell (Ni , Cj ), let ξ (Ni , Cj ) represent the set of patterns that realize the cell and |ξ (Ni , Cj )| the number of patterns in that set. There will be at least one inflectional pattern associated with every cell, since the existence of the pattern is a precondition for recognizing the cell in the first place. In every class system, there will also be a maximum number of patterns realized by any cell, though there need not be a unique ‘maximally allomorphic’ cell.

Table 5.2 Schematic paradigm structure

C1 C2 C3 C4

N1

N2

(N1 ,C1 ) (N1 ,C2 ) (N1 ,C3 ) (N1 ,C4 )

(N2 ,C1 ) (N2 ,C2 ) (N2 ,C3 ) (N2 ,C4 )

96

james p. blevins

The space of possible inflection classes can then be defined with reference to the allomorphic variation exhibited by individual cells. It is usually assumed that two inflection classes can be distinguished if they exhibit (morphologically conditioned) allomorphy in any cell. From this assumption, it follows that the maximally allomorphic cell defines the minimal number of classes. This minimum just reflects the fact that there must be at least as many classes as there are realizations for the maximally allomorphic cells, since otherwise some class would have multiple realizations for those cells. At the other extreme, the largest space of classes corresponds to the cell product, that is, the product of all of the realizations of the individual cells. In the case of the system in Table 5.2, the cell product is defined as |ξ (N1 , C1 )| × |ξ (N1 , C2 )|, . . . , ×|ξ (N2 , C4 )|. The cell product defines the largest class system, since it exhaustively enumerates all of the combinations of patterns exhibited by individual cells. Continuing with the schematic system in Table 5.2, let D represent the system and CD be the number of inflection classes in D. The assumptions adopted so far dictate only that the value of CD must fall somewhere between the number of realizations associated with the maximally allomorphic cell and the cell product. However, as Carstairs (1983) remarks, inflection class systems are not normally distributed within this space. Instead, they appear to cluster closely around the number of realizations associated with the maximally allomorphic cell: when we apply the traditional notion of ‘paradigm’, we find that the actual total of paradigms is at or close to the minimum logically possible with the inflexional resources involved, and nowhere approaches the logical maximum. (Carstairs 1983: 127)

To account for this clustering, Carstairs (1983: 127) proposes the PEP as ‘an absolute constraint on the organization of the inflexional resources for every word-class in every language’ which has the effect of ‘keeping the total of paradigms for any word-class close to the logical minimum’. The PEP thus requires that the class size of any wordclass system is equal to (‘or close to’) the number of realizations associated with the maximally allomorphic cell. In principle, the PEP would appear to impose a maximally restrictive constraint on class size, since it effectively requires that each inflectional system must be organized into the smallest possible set of classes. In practice, the restrictiveness of the PEP depends on how close is ‘close enough’ for compliance and, more fundamentally, on how classes are individuated. One immediate qualification is introduced by the decision to ‘disregard . . . stem alternations’ (Carstairs 1983: 120) and restrict the notion of ‘inflectional resources’ relevant to the PEP to ‘affixal exponents’. This exclusion reduces the overall number of inflection classes in languages with distinctive, class-specific stem alternations, and contributes to the goal of bringing the count closer to the logical minimum. Yet even as a constraint solely on the distribution of affixal resources, the PEP places an extremely tight constraint on the relationship between cells in a paradigm. In a

inflectional paradigms

97

system that conforms to the PEP, each maximally allomorphic cell partitions the system into classes, and every pair of maximally allomorphic cells partition the system into the same classes. No cell in a paradigm can have realizations that vary independently of the realizations of any maximally allomorphic cell. This means that for every realization of every maximally allomorphic cell there will be a unique realization in every other cell (including other maximally allomorphic cells). This entails that every maximally allomorphic cell is diagnostic of class. Moreover, since every class contains a maximally diagnostic cell, it follows that every class will be identifiable from the realization of a single cell. So a system that conforms to the PEP will realize the pedagogical ideal in which a single principal part suffices to identify class.

5.3.2 Gender-driven Economy in German There are prima facie counterexamples to any principle that imposes this kind of tight organization on paradigm structure. As Carstairs acknowledges, traditional descriptions of German noun declensions do not appear to conform to strictures of the PEP. Table 5.3 lists the exponents that Carstairs isolates from the traditional principal parts of these declensions.7 Since the smallest class space is defined by the most maximally allomorphic cell, the five exponents that realize the non-dative plural in Table 5.3 define a minimum class size of five. This corresponds closely to the number of plural classes recognized in traditional sources such as Duden (2005) (provided that stem alternations are again disregarded). The five basic plural classes P1–P5 are illustrated in Table 5.4, in which each pattern is classified by the shape of its plural ending and—for completeness—whether it is, or can be, based on an umlauted variant of the singular stem. A comparison of the forms in the first two blocks of rows shows that masculine and neuter nouns both follow all five patterns, although comparatively fewer neuter nouns follow the ‘weak’ pattern P2 and only a very few exhibit umlaut in P3. The bottom block of rows shows that feminine nouns do not form plurals in -er and that they tend to require umlauted stems in patterns P3 and P5. Table 5.3 Exponents of principal parts in German Gen Sg Nom/Acc/Gen Pl

Ø, -(e)s, -(e)n, -(e)ns Ø∼-e, -¨(e), -er∼-¨er, -s, -(e)n

Source: Carstairs (1983: 125).

7 Parentheses and dashes mark alternations treated as phonologically conditioned or stylistic, and exponents of the form ‘-¨er’ indicate an umlauted stem vowel.

98

james p. blevins

Table 5.4 Plural patterns in German P1

P2

P3

P4

P5

Ending Stem

-s –uml

-(e)n –uml

-e –uml

-e +uml

-er +uml

Ø –uml

Ø +uml

N/A/G Dat (Masc)

Uhus Uhus ‘owls’

Prinzen Prinzen ‘princes’

Hunde Hunden ‘dogs’

B¨unde B¨unden ‘waistbands’

M¨under M¨undern ‘mouths’

Balken Balken ‘beams’

G¨arten G¨arten ‘gardens’

N/A/G Dat (Neut)

Autos Autos ‘cars’

Ohren Ohren ‘ears’

Jahre Jahren ‘years’

Fl¨oße Fl¨oßen ‘rafts’

L¨ander L¨andern ‘countries’

Muster Mustern ‘patterns’

Kl¨oster Kl¨ostern ‘cloisters’

N/A/G Dat (Fem)

Bars Bars ‘bars’

Regeln Regeln ‘rules’

— — —

H¨ande H¨anden ‘hands’

— — —

— — —

T¨ochter T¨ochtern ‘daughters’

Source: cf. Duden (2005: 226).

Table 5.5 Singular declensional patterns in German Masc

Nom Acc Dat Gen

S1

S2

Neut S1

Fem S3

Pegel Pegel Pegel Pegels ‘level’

Prinz Prinzen Prinzen Prinzen ‘prince’

Segel Segel Segel Segels ‘sail’

Regel Regel Regel Regel ‘rule’

Source: cf. Duden (2005: 197).

To conform to the PEP, the choice of plural ending must determine the regular singular patterns, illustrated in Table 5.5.8 Pattern S3 characterizes feminine nouns, which are invariant in the singular. Neuters likewise follow just the ‘strong’ pattern S1, in which the genitive singular is marked by -(e)s. Masculine nouns exhibit the only variation, as they may follow either the strong pattern S1, marked by -(e)s or the weak pattern S2, marked by -(e)n. It is sometimes claimed for German that ‘the choice of plural formation depends largely on gender and/or inflection class as manifested . . . in the singular’ (Laaha et al. 8 This table omits the proper name class listed in Duden (2005: 197), which patterns with S1 nouns. It also excludes the pattern exhibited by name ‘name’ and a small handful of other nouns, which have accusatives and datives in -en and genitives in -ens.

inflectional paradigms

99

Table 5.6 Combinations of singular and plural patterns in German uml P1 P2 P3 P3 P4 P5 P5

– – – + + – +

S1 Neut

S2 Masc

S3

Masc uhu staat hund bund mund balken garten

auto ohr jahr (floß) land muster (kloster)

— prinz — — — — —

kamera regel — hand — — tochter

Fem

Table 5.7 Co-occurrence of plural and genitive singular endings in German S1 Pl P1 P2 P3 P4 P5

-s -(e)n -e -er Ø

S2

S3

Masc

Neut

Masc

Fem

-s -s -s -s -s

-s -s -s -s -s

— -(e)n — — —

Ø Ø Ø — Ø

2006: 279). This claim is unfounded, as Table 5.6 shows. Matching the singular patterns in Table 5.5 with the plural patterns in Table 5.4 defines nineteen combinations, which highlight the dissociation of plural from singular patterns. The affixal variation is isolated in Table 5.7, which plots the co-occurrence of plural and genitive singular endings.9 These patterns are obtained from Table 5.6 by collapsing pairs of rows that differ solely in stem umlaut, and replacing the exemplary lexemes by plural and genitive endings. At first glance, this pattern appears incompatible with the PEP, since every plural ending except -er in the second column corresponds to two or even three distinct genitive singular endings. Yet on closer inspection the system is in fact almost maximally economical. As Carstairs argues, traditional ‘declensions’ (and ‘conjugations’) can be brought into closer conformance with the PEP by restricting attention to genuine inflectional variants, and grouping paradigms into common classes if they differ 9

Though omitting -(e)ns, which occurs with a small and declining set of nouns.

100

james p. blevins Table 5.8 Mildly uneconomical class space in German

Masc Neut Fem

P1

P2

P3

P4

P5

S1 S1 S3

S1 S2 S1 S3

S1 S1 S3

S1 S1 —

S1 S1 S3

solely with respect to lexically-conditioned properties. In the specific case of German, Carstairs (1983: 126) observes that ‘Gender is lexically determined for German nouns; and we can readily combine each of the Feminine-only “declensions” . . . with some non-Feminine “declension”, just as we traditionally combine Latin Neuters and nonNeuters in Declension II’. Returning to the classes defined by the P1, P3, and P5 plural endings in Table 5.7, one can see that they exhibit a general contrast between masculines and neuters, which are marked by -s in the genitive singular, and feminines, which are uninflected in the singular. The distribution of umlauted vowels (suppressed in Table 5.7) is likewise a lexical property of nouns in classes P3 and P5. The strongest challenge to the PEP is posed by class P2. There is, first of all, some artificiality in collapsing what are arguably two distinct plural strategies. Plurals in -(e)n are the default for ‘native’ feminine nouns, as well as for feminine nouns formed with productive endings such as -ung and -heit, as expressed in the second ‘basic rule’ (Grundregel) in Duden (2005: 193). The formation of masculine and neuter plurals in -(e)n is much more restricted and cannot be regarded as productive in the modern language.10 Even within the non-productive subclass, neuters have unambiguous singular forms. So it is just the masculines that show inflectional variation, with weak nouns such as prinz following the weak singular pattern S2, and mixed nouns like staat following the strong singular pattern S1. The resulting class space is exhibited in Table 5.8. Hence by consolidating gender-conditioned variation, one can arrive at a space of six classes, assuming just two P2 subclasses, or seven classes, if productive feminines in -(e)n are distinguished from the masculines and neuters. This exceeds, but is indisputably close to, the minimum of five determined by the patterns of plural exponence. One can in principle bring the system into conformance with the PEP by coercing plurals in -(e)n into a single class and making a ‘specific exemption’ for mixed paradigms, as Carstairs (1983: 127) suggests. This kind of exception is of course very different from the consolidation of lexically-conditioned variation. However, it is also true that the evaluation of a general principle like the PEP should not hinge too directly on patterns that even traditional sources regard as mixed or hybridized. 10 There is also psychological and neurolinguistic evidence for distinguishing feminines in -(e)n from the corresponding masculines and neuters (Clahsen 1999; Penke et al. 1999).

inflectional paradigms

101

It is arguably more instructive to diagnose the general characteristics of German that allow the PEP to work, or at least to work as well as it does. One pivotal property is the clean dissociation of inflectional and lexical variation. The patterns of plural affixation P1–P5 can reasonably be regarded as inflectional, while affixal variation between the singular patterns S1 and S3 can, as Carstairs argues, be attributed to lexical factors. Hence there is only one case, involving masculines of class P2, in which there is inflectional variation in both the plural and singular. It is precisely this case that is problematic for the PEP. Although the separation of lexical and inflectional variation promotes economical paradigm organization in German, this separation is less a feature of inflection class systems in general than a symptom of the nearcomplete loss of singular contrasts in German. Hence a useful test case for the PEP comes from declensional systems that exhibit inflectional contrasts in the singular and plural.

5.3.3 Uneconomical Partitives in Finnish One immediate consequence of the relation between paradigm economy and predictability is that the PEP will tend to be violated by any language whose description requires more than one principal part, since multiple parts imply that the realization of multiple cells must vary independently. Finnish provides a useful test case, given that descriptions of declensions often list up to five principal parts (typically the nominative, genitive, and partitive singular, along with the partitive plural and a plural ‘local’ case such as the inessive). Some of these forms identify stem alternations that are disregarded in determining compliance with the PEP, while others identify affixal patterns. Since the local case forms have mostly invariant endings, the grammatical case exponents in Table 5.9 exhibit nearly all of the inflectional variation that is relevant to the PEP. Examination of Table 5.9 identifies the partitive singular as the most highly differentiated cell, with three distinct patterns -A, -tA, and -ttA. The archiphoneme ‘A’ ranges over the vowels a and a¨ , which show harmony with the final stem vowel, so that the

Table 5.9 Grammatical case exponents in Finnish

Nom Gen Part

sg

pl

Ø -n -A, -tA, -ttA

-t -den∼-tten, -en -A, -tA

Source: Karlsson (1999).

102

james p. blevins Table 5.10 Grammatical case forms in Finnish Part Sg

Part Pl

Part Sg

Part Pl

(Gen Pl)

A A tA tA ttA

A tA A tA tA

asemaa perunaa lohta leikkuuta huonetta

asemia perunoita lohia leikkuita huoneita

(asemien) (perunoiden) (lohien) (leikkuiden) (huoneiden)

‘position’ (13) [15] ‘potato’ (17) [3] ‘salmon’ (33) [24] ‘haircut’ (25) [18] ‘room’ (78) [4]

A

A∼tA

karitsaa

A

lahtea lahta

(karitsojen) (karitsoiden) (lahtiin)

‘lamb’ (15) [6]

A∼tA

karitsoja karitsoita lahtia

‘bay’ (34) [12]

Source: Pihel and Pikam¨ae (1999).

three patterns in Table 5.9 have six surface realizations. The partitive plural is realized by the first two endings, -A and -tA, but these endings are usually described as varying independently of the partitive singular endings. The genitive plural is realized by three endings, but two of these, -den and -tten are regarded as variants.11 The grammatical case cells in Finnish thus have at most three realizations, and determine a cell product of 12 (3 × 2 × 2). The variation exhibited by grammatical case forms (according to standard descriptions) is set out in Table 5.10.12 The first two columns in Table 5.10 plot the co-occurrence of partitive singular and plural exponents, followed in the next two columns by forms of exemplary items (including predictable genitive plurals). We can with no loss of generality restrict attention to word types with unique partitive singular and plural affixes, exhibited in the first five rows in Table 5.10. These word types define the five classes C1–C5 in Table 5.11. A class size of five is closer to the maximum of six defined by the product of the partitive exponents than to the minimum of three dictated by the PEP. But it is still far from the cell product of 12, due to the fact that genitive plural endings do not add any new classes. Instead, variation in the genitive plural largely conforms to the expectations of the PEP in aligning with the partitive plural, with -A implying -en and -tA predicting -den (Karlsson 1999: 92f.). The challenge presented by Finnish is that the partitive plural is itself not predictable from the partitive singular. The four logical possibilities defined by the exponents -A, and -tA are illustrated in the first four rows in Table 5.10: singular -A co-occurs 11

‘The ending -den can always be replaced by the ending -tten’ (Karlsson 1999: 93). The numbers in parentheses identify the word type number from the authoritative Nykysuomen sanakirja ‘Dictionary of modern Finnish’ (H¨akkinen 1990), which are used in Pihel and Pikam¨ae (1999), and the number in square brackets indicates how many of the 85 word types exhibit that pattern of partitive exponence. 12

inflectional paradigms

103

Table 5.11 Classes defined by unique partitive exponents

Part Sg Part Pl Gen Pl

C1

C2

C3

C4

C5

A A en

A tA den

tA A en

tA tA den

tA ttA den

with each of the plural endings -A and -tA and the same is true of singular -tA. Since there is no grammatical gender in Finnish, it is not possible to collapse the classes in Table 5.11 into ‘macroparadigms’ whose internal variation is conditioned by gender differences, as in the previous case of German. Nor are there any other morphosyntactic properties that condition the variation in partitive exponents. It is also not plausible to maintain the PEP by treating the patterns in Table 5.10 as conditioned by variation in the shape or prosody of stems. Instead, the independent variation of partitive singular and plural exponents define a product class space that violates the PEP. As the discussions of German and Finnish clarify, the PEP is ultimately a constraint on the form of morphological descriptions rather than on the organization of the systems themselves. This is to some degree unavoidable, given that inflection classes are part of a descriptive frame of reference that is imposed on languages rather than ‘discovered’ in them. Questions about the number and type of inflection classes in a language cannot even be formulated without specifying the purposes for which the classes are defined. The status of classes is obscured to some degree by the fact that the agreement regarding class individuation in traditional sources derives from common descriptive or pedagogical goals. However, this consensus does not rest on any general task-independent criteria for distinguishing classes. Two paradigms that exhibit distinct inflectional patterns in each of their cells would be assigned to distinct classes in any system of classification. But how many cells must vary? Does one always suffice, or does partial overlap between two paradigms give rise to ‘subclasses’? Does it matter if variation involves suppletion or some other form of irregularity? How many items must follow an inflectional pattern in order to constitute a class? Presumably single suppletive items do not form classes. But what then is the item threshold that separates classes from residual suppletive patterns? The issues raised by these questions are not addressed by dictating a consistent policy governing the definition of classes, subclasses, item-specific patterns, etc. The true challenge lies in arriving at a principled, task-independent, basis for this definition. The fact that this challenge is seldom if ever confronted reflects an understanding that there is not some fixed number of classes to be discovered in a language, but rather that the number assigned depends on the level of precision to which the classes are described

104

james p. blevins

or the purposes for which they are defined. The following quotation describes the situation in Finnish; the status of Estonian is comparable (Blevins 2007, 2008). There is no consensus on how many inflectional classes there are for nominals and verbs. Traditional Finnish lexicography as manifested in Nykysuomen sanakirja (Dictionary of modern Finnish, 1951–1961) postulates 82 inflectional classes for nominals, whereas at the other extreme, a generative description such as Wiik (1967) operates with none but a wealth of ordered (morpho)phonological rules. A surface-oriented morphological approach would recognize at least 10 nominal inflectional classes. (Karlsson 2006: 476)

This indeterminacy highlights limitations of economy conditions which impose a numerical bound on constructs, in this case ‘inflection classes’, that are part of the description of a morphological system. The most direct development of the PEP is the No-Blur Constraint (Carstairs-McCarthy 1994), which expresses economy in terms of restrictions on the relation between affixal resources and classes. This alternative thus inherits any limitations that accrue from a purely affixal measure or from assumptions about the determinacy of inflection classes.13 Nevertheless, these issues mainly pose challenges to the specific formulation of PEP and successor constraints. Constraints that attempt to characterize the highly economical and cohesive organization of inflectional paradigms draw attention to a pattern which ‘by itself is enough to establish the paradigm as an important concept within linguistic theory’ (Carstairs 1983: 127). The observation underlying the PEP also cannot be dismissed as a descriptive artifact. The ‘tendency’ that Carstairs (1983: 127) discerns ‘towards keeping the total of paradigms for any word class close to the logical minimum’ seems real enough, even if it does not appear to be susceptible to formalization directly in terms of numerical bounds on class size or affixal classification.

5.4 Modelling Paradigmatic Structure

.............................................................................................................................................................................

From a classic WP perspective, paradigm economy effects do not derive from constraints that regulate the economical organization of paradigms or the ‘inflectional resources’ distributed across their cells. Rather, these effects are taken to be secondary reflexes of the implicational structure of an inflectional system. This structure is not explicitly formalized in the classic WP tradition but merely exhibited by interdependencies between forms of exemplary items. The implicative Paradigm Structure Conditions proposed by Wurzel (1984) in the context of a model of Natural Morphology thus provide one of the first explicit representations of the structure of inflectional paradigms. 13

Some of the empirical consequences of these assumptions are discussed in Stump (2005b).

inflectional paradigms

105

Observation of complicated paradigms shows that implicative relations do not only obtain between one basic inflexional form, either lexical (sobaka [Russian ‘dog’ JPB]) or non-lexical (M¨anner [German ‘men’ JPB]), and all the other inflexional forms, but exist throughout the whole paradigm: all paradigms (apart from suppletive cases) are structured on the basis of implicative patterns which go beyond the individual word, patterns of varying complexity. Of particular complexity in this respect is, for example, the implicative pattern of the i-declension in Latin: /im/ in the a.sg. ⊃ /¯ı/ in the abl.sg. ⊃ /¯ıs/ in the a.pl. ⊃ /ium/ in the g.pl. ⊃ . . . From the a.sg. form we can derive, via a number of steps, all other forms, but not vice-versa. Since the implicative patterns determine the structure of the paradigms of a language we call them implicative paradigm structure conditions (PSCs). (Wurzel 1984: 208)

Yet the use of the material conditional ‘⊃’ to model implicational relations constrains the descriptive scope of PSCs, restricting their application to exceptionless patterns. Some patterns are exceptionless, as in the case of the relation between Finnish partitive singulars and genitive plurals in Table 5.10. However, in many if not most patterns, one cell or form is only partially informative about another, and a general analysis of implicational structure must be able to represent this partial information, and the ways in which it can be combined.

5.4.1 Implication as Uncertainty Reduction Standard notions from information theory permit a much more flexible representation of implicational patterns. The basic idea is to think of morphological variation first of all in terms of uncertainty; the more variants of a cell there are (and the more uniform their distribution), the harder it is to guess which variant realizes a particular cell. Then implicational dependencies correspond directly to uncertainty reduction; there is an implicational relation between one cell and another to the extent that knowing the realization of the first cell reduces uncertainty about the realization of the second. Unlike logical implication, this is not an all or nothing relation. Knowledge of one cell may merely reduce the uncertainty of another, without eliminating it altogether. For example, as Table 5.6 indicates, knowing that a German noun has the plural ending -(e)n implies that it is either a feminine noun that follows the (invariant) singular pattern S3 or a masculine noun that inflects according to singular pattern S2. These intuitive notions have very straightforward formalizations. To begin with, the (self-)information carried by an exponent x can be expressed in terms of surprisal, as in (1). The informativeness of the exponent, Ix , corresponds to the negative log of its probability p(x), which just reflects the fact that the less likely x is, the more information it conveys when it does occur. (1)

Ix = − log2 p(x)

106

james p. blevins

The uncertainty or entropy (Shannon 1948) of a cell is then defined by averaging the surprisal values of the exponents that realize the cell. The entropy of a cell is defined in (2) in terms of a standard entropy definition by treating a cell C as a random variable that takes as values a set X of exponents x1 . . . xn . ∑ (2) H(C) = − p(x) log2 p(x) x∈X

The surprisal value of an exponent cannot be estimated without information about its distribution, and the same information is required for an accurate measure of cell entropy. However, even without distributional information, a worst-case estimation of cell entropy can be obtained by treating the patterns of exponence as equiprobable. This reflects the fact that the uncertainty measured by entropy increases as a function of the number of alternatives and the uniformity of their distribution (since both factors make it more difficult to guess the value). In this limiting case, the probabilities ‘cancel out’ and entropy reduces to the log2 of the number of alternatives in the set X = x1 . . . xn , as in (3). (3) H(C) = log2 (n) For the purposes of linguistic analysis, an extremely useful property of entropy is that estimations can be obtained from a range of descriptive sources. Entropy ceilings can be obtained for languages that are only described by traditional grammars, provided that the grammars contain a complete description of form inventories. The more balanced and detailed the corpus resources available for a language, the more accurate the corresponding entropy measures. For a simple illustration, consider the Finnish grammatical case exponents, repeated in Table 5.12. The partitive singular is a maximally allomorphic cell, with three exponents. In the example in Table 5.12, the worst-case entropy of the partitive singular cell is log2 (3), or approximately 1.58 bits. This value assumes that the exponents are equally probable; any distributional skewing that makes any exponent more (or less) likely than the others will lower the entropy. Entropy represents the uncertainty associated with a cell or paradigm. Once an uncertainty estimation has been obtained, implicational structure can be modelled by uncertainty reduction. Uncertainty reduction is expressed by conditional entropy,

Table 5.12 Grammatical case exponents in Finnish

Nom Gen Part

sg

pl

{Ø} {-n} {-A, -tA, -ttA}

{-t} {-den∼-tten, -en} {-A, -tA}

inflectional paradigms

107

H(C2 |C1 ), which measures the amount of uncertainty that remains about C2 given knowledge of C1 . This relation can be defined as in (4). (4)

H(C2 |C1 ) = −

∑ x∈X

p(x)

∑

p(y|x) log2 p(y|x)

y∈Y

As (2), the cells C1 and C2 are random variables that take as values the sets of exponents X = x1 . . . xn and Y = y1 . . . ym . The only new element in (4) is p(y|x), the conditional

probability of an exponent y given an exponent x.

5.4.2 Diagnosticity and Economy These basic notions of uncertainty and uncertainty reduction permit a straightforward representation of the implicational structure of a morphological system. The more information that a cell C1 provides about another cell C2 , the lower the conditional entropy value H(C2 |C1 ) will be. If knowledge of C1 completely identifies the value of C2 , then H(C2 |C1 ) will approach 0. If C1 provides no information about C2 , then H(C2 |C1 ) will preserve the uncertainty of H(C2 ). The selection of principal parts can be guided by entropy reduction. A set of cells function as principal parts within a word class whenever, for any paradigm in that class, knowledge of those cells reduces uncertainty about any other cells in the paradigm. There need not be a unique diagnostic cell or set of cells. The availability of a range of different diagnostic combinations clearly enhances the robustness of form deduction, since speakers are not dependent on encountering one uniquely informative form of a paradigm. The informal idea that diagnosticity correlates with variability across word or inflection classes can be expressed more precisely in terms of the uncertainty reduction measured by the conditional entropy. From the present perspective, one can see that the choice of leading forms is to a large degree arbitrary. A pedagogical or reference grammar might use seemingly arbitrary criteria to select a particular cell or cell set. A description might select the smallest set of cells, the set with the most highly frequent members, or, more capriciously, the cell or cell set with the morphologically simplest members, etc. Since any fully diagnostic set of cells will do, all are equally suitable and the arbitrariness involved in selecting one is harmless.14 An uncertainty-based approach also offers a useful perspective on the central insight expressed by the PEP, namely that the variation exhibited by maximally allomorphic cells constrains the class-defining distribution of variants in other cells. Taking the number of variants in the maximally allomorphic cell to define class size (or distribution of variants) determines a tight constraint on the distribution of allomorphs—too tight as the example of Finnish suggests. 14 There is also a trade-off between the number of cells required to identify class in a system and whether one uses the same cells to identify class, as Finkel and Stump (2007, 2009) show.

108

james p. blevins

However, the same fundamental intuition can be expressed by means of a measure that defines economy more directly in terms of relations between maximally allomorphic cells and other cells. An economy measure based on the interdependence of cells neither requires an independent consensus about the number of classes in a language nor methods for imposing minimized descriptions, since interdependencies tend to be preserved irrespective of the number of classes or the ‘granularity’ of the classification. To some degree the central role of cell interdependence is obscured in the paradigm economy literature by the decision to view inflectional resources in terms of separate inventories of stems and inflections, given that the disassembly of forms disrupts implicational patterns based on whole words (termed ‘gestalt exponence’ in Ackerman et al. 2009). Maximally allomorphic cells are significant within a system because they determine bounds on variation within the system. By definition, a maximally allomorphic cell is associated with the largest number of variants. Another way of putting this is that maximally allomorphic cells exhibit the greatest ‘morphological uncertainty’; it is hardest to guess the realizations of these cells, because they are associated with the largest number of choices. The variation in a maximally allomorphic cell also serves a diagnostic function within a system. Either the variation in such a cell correlates with the variation in other cells, or else the system consists of mutually independent cells. This relation can again be phrased in terms of uncertainty. Knowing the realization of a maximally allomorphic cell reduces uncertainty about the realization of at least some other cells, except in the case where the cells of the system are mutually independent. To express these relations schematically, let CM be a maximally allomorphic cell and H(CM ) represent the uncertainty associated with CM . For a pair of cells C and D, let H(C|D) represent the uncertainty that remains about C given knowledge of D. Then a paradigm consisting of cells C1 , . . . , Cn will exhibit the minimal interdependence assumed in a classic WP model if H(Ci ), the average uncertainty associated with the cells, falls within the range specified in (5). (5) H(Ci |CM ) < H(Ci ) ≤ H(CM ) The bounds associated with maximally allomorphic cells in (5) merely ensure that a system does not consist of mutually independent cells. The average uncertainty H(Ci ) is required to be greater than the average uncertainty given knowledge of a maximally allomorphic cell, H(Ci |CM ), and no greater than the uncertainty of the maximally allomorphic cell, H(CM ), itself. Economy principles impose further restrictions within these limits. The PEP requires in effect that knowledge of a maximally allomorphic cell reduces all uncertainty, that is, that H(Ci |CM ) approaches zero. This follows from the assumption that the realization of a maximally allomorphic cell identifies the class of an item, together with the assumption that class determines the realizations of the other cells. The formulation of economy in terms of uncertainty clarifies two issues. The first is that classes function essentially as proxies for patterns of interdependency in the PEP. A constraint that stipulates that maximally allomorphic cells eliminate uncertainty in

inflectional paradigms

109

other cells achieves the effect of the PEP without any reference to classes. The second issue is that the focus on maximally allomorphic cells exhibits the same bias as traditional principal part analyses. A system that conforms to the PEP is one in which a single diagnostic form from a maximally allomorphic cell determines all of the forms of a non-exemplary item. In much the same way that principal part analyses tend to ignore the informativeness of non-diagnostic forms, the PEP defines economy solely in terms of the correlation between maximally allomorphic cells and patterns of interdependency. From the perspective of uncertainty reduction, there is no reason to privilege maximally allomorphic cells in this way, since the interdependency between other cells will also determine the cohesion and economy of a system. The idea that knowledge of some forms of an item reduces uncertainty about other forms of the item remains implicit in classic WP models. The first systematic attempt to measure this effect is proposed in the context of the Low Entropy Conjecture of Ackerman and Malouf (2013), which suggests that the average uncertainty of a cell given knowledge of another is not only lower than the uncertainty of the cell in isolation, but falls within a fairly narrow range.

5.5 Conclusion

.............................................................................................................................................................................

This chapter has devoted considerable attention to the role played by implicational relations in inflectional paradigms. In part this focus reflects the fact that the implicational structure of inflectional paradigms is their most distinctive characteristic and also provides some of the strongest evidence for recognizing paradigms as fundamental units of lexical organization. Although implicational relations appear to be central to the organization of language in general, their effects are particularly salient in the domain of inflection. It is the fact that inflection operates over a closed, relatively uniform space that makes inflectional processes highly predictable, and in turn contributes to their productivity. Any speaker who knows that auto is a regular count noun in English can deduce that it has a 2-cell paradigm defined by the number contrast between singular auto and plural autos.15 The two forms also predict each other, given that the regular plural marker is phonologically conditioned in English. From the fact that German Auto is neuter and ends in a vowel that does not conform to native nominal phonotactics, a German speaker can deduce that the plural forms and genitive singular of the noun auto (classes P1/S1 in Table 5.6) are realized by Autos, while all other singular cells are realized by Auto. The twenty-two-cell paradigm of the Finnish counterpart auto (class 1, represented by asema in Table 5.10) is likewise defined by the two numbers and eleven regular cases in Finnish. The interpredictability of forms 15 Though it does not affect the point at hand, the ‘Anglo-Saxon’ genitive marker ’s is almost certainly best regarded as a phrasal affix (Hockett 1947: 242; Anderson 1992: 212; Blevins 2006a), not as a lexical case exponent, as suggested in Zwicky (1987) and Huddleston and Pullum (2002).

110

james p. blevins

is more complex in Finnish, due to larger form inventories, consonant gradation and other sources of variation. Yet, as in English, interpredictability is facilitated by the uniformity of declensional paradigms. A speaker encountering a new noun in Finnish can assume that it has twenty-two case-number forms, and even deduce the shape of those forms with levels of confidence that depend on the morphological information expressed by the form they have encountered.16 Contributing to the regularity of inflection is the fact that each cell C in a feature space is related to other cells along two dimensions. On the one hand, C is related to other cells in the same paradigm, that is, to cells that have the same intrinsic features as C. On the other hand, C is related to counterpart cells in other paradigms that have different intrinsic features from C but contain the same (or similar) paradigmatic features. The uniform organization of inflectional paradigms contrasts starkly with the more variable structure exhibited by ‘families’ of derivationally-related forms. Derivational processes are traditionally classified as ‘exocentric’ in the sense that they define new items by changing the word class, valence, or other intrinsic properties of the items to which they apply. Accounts that attempt to demarcate inflection from derivation often focus on the fact that such ‘property-changing’ processes are non-monotonic. But from the present perspective, it is at least equally important that they also do not define a finite set of forms within a uniform feature space. From the word class of an item, one cannot in general predict the number and type of derivational formations in which it occurs. Given a list of derivational processes active in a language, it is of course possible to assign a uniform family of ‘potential’ forms to all of the members of a word class. Yet the uniformity achieved is deceptive, because it collapses a critical distinction between those forms that are established in a language and those that are merely possible in principle.17 Defining expanded derivational paradigms does not make it any easier for a speaker to predict the derivational formations that are attested and in use within a language. The contrast with compounds is even clearer. Of the infinitely many possible noun compounds in English only a comparatively small number are established, and a speaker cannot predict the set of established compounds containing an item from the item itself. The lack of predictability is due in large part to the high degree of item-specific variation exhibited by derivational families, which does not give rise to form classes that sanction class-specific inferences. The contrast between the cohesive implicational structure of an inflectional paradigm and looser confederations of derivational forms are also reflected in the ways that morphological systems are described and analysed. It is rare for an inflected form to be described as established on its own. The availability of a form tends to correlate instead with the productivity of a whole pattern. Conversely, notions like 16

There may also be some residual comitative, abessive, or instructive forms. The available psycholinguistic evidence suggests that established forms, not potential forms, are psychologically relevant. A range of experimental studies have found that the processing of items in Germanic, Finnic, and Semitic, among other languages, is facilitated by the semantically transparent items that make up its morphological family and inhibited by semantically opaque items (Schreuder and Baayen 1997; de Jong 2002; Moscoso del Prado Mart´ın et al. 2004a, 2004b). 17

inflectional paradigms

111

‘morphological gap’, ‘suppletion’, and even, to a large degree, ‘syncretism’ are mainly or exclusively applied to inflectional paradigms. Derivational families do not usually have ‘gaps’ because they do not define implicational relations over a closed and uniform space that give rise to expectations about the existence of specific forms, even where the shape would be predictable. Suppletion can likewise only arise where there are definite assumptions about the shape of particular members of a form set. Whereas inflectional paradigms generate strong assumptions of this nature, derivational families do not. Syncretism presupposes a similar structure, as syncretic patterns imply the existence of independent cells that can be associated with fully or partially identical forms. But the notion of a derivational cell is not defined in the absence of a set of features that specify a morphosyntactic grid within which to place the cells. More generally, descriptions of inflectional systems tend to assume a closed and uniform feature space that sanctions inferences about the existence and shape of forms. In many languages, there is evidence that this space is further partitioned into units that correspond to the traditional notion of a paradigm and that these units serve as the locus of form deduction and economy effects.

chapter 6 ........................................................................................................

INFLECTION CLASSES ........................................................................................................

gregory stump

6.1 Introduction

.............................................................................................................................................................................

Many are the students who, in Latin 101, have asked themselves why the inflection of Latin nouns has to be complicated by declensional distinctions. Whereas a noun’s case and number distinctions generally have semantic or syntactic significance, declensionclass distinctions have neither; for instance, although the declensional system of Latin encompasses both adjectives and nouns, an adjective modifying a Latin noun agrees with it in case and number, but not in its declension-class membership1 (e.g. r¯ex bonus ‘good king’: bonus matches the singular number and nominative case of r¯ex, but r¯ex belongs to the Third declension while bonus belongs to the Second). It is true that Latin declensions correlate in certain ways with genders, which are syntactically if not semantically significant (thus, bonus matches the masculine gender of r¯ex), but this correlation is imperfect; for instance, although First-declension nouns are mainly feminine, they do include certain masculines, for example po¯eta ‘poet’. Moreover, nouns belonging to the same gender may differ in declension; for instance, po¯eta, am¯ıcus ‘friend’, and r¯ex are all masculine, but po¯eta belongs to the First declension, am¯ıcus to the Second, and r¯ex to the Third. The naïve conclusion is therefore that declensional distinctions (and inflection-class distinctions generally) would not exist in an ideal language. This innocent conclusion does not alter the fact that in real languages, inflectionclass distinctions are extremely common. In the terminology of Aronoff (1994), such distinctions are morphomic: they are visible to a language’s morphology, but not to its other grammatical components; no rule of phonology or syntax or semantics is sensitive to them. As a consequence, one should not expect them to exhibit the same kinds of usefulness as morphosyntactic properties such as case, number, and gender. 1 An adjective’s own declension-class membership may, of course, accidentally coincide with that of the noun it modifies.

114

gregory stump

But they are useful in their own way. A lexeme’s inflection-class membership is part of its lexical identity, and may therefore serve a disambiguating function; for instance, the lie1 in He just lay at home inflects as a strong verb, while the lie2 in He just lied at home inflects as a weak verb. Even in the absence of ambiguity, morphological expressions of inflection-class membership are as distinctive a part of a lexeme’s identity as the phonology of its stem(s). In this chapter, I discuss a number of fundamental issues concerning the nature of inflection classes.2 In Section 6.2, I give a formal characterization of the notion ‘inflection class’ and exemplify this notion with evidence from Icelandic. In Section 6.3, I discuss the notion ‘canonical inflection class’ proposed by Corbett (2009b); I examine a range of simple and complex deviations from the canonical ideal in Section 6.4. In Section 6.5, I discuss the diverse ways in which a lexeme’s inflection-class membership may be expressed morphologically. In Section 6.6, I consider the different kinds of extra-morphological correlates that a language’s inflection classes may exhibit. In Section 6.7, I examine the diachronic development of inflection classes: how new inflection-class distinctions arise, how a lexeme may change its inflection-class membership, and how inflection-class distinctions may be eliminated. In Section 6.8, I discuss the appropriate mode of representation for inflection classes in the formal definition of a language’s grammar and the status of inflection classes as a dimension of typological variation. I conclude in Section 6.9 with a summary of factors that contribute to the ongoing challenge that inflection classes pose for linguistic theory and analysis.

6.2 What is an Inf lection Class?

.............................................................................................................................................................................

In a given language, a lexeme L generally has a set M of morphosyntactic property sets with whose members it may be associated in syntax, and the inflectional morphology of the language provides a set of inflectional markings that allow each association of L with a member of M to be realized morphologically. Under these circumstances, I shall call the complete set of pairs of the form ⟨L, ⟩ (where  is a member of M) the content paradigm of L, and its member pairings I shall call cells; correspondingly, the set containing the morphological realization of each cell in L’s content paradigm is its form paradigm. In the simplest cases, lexemes belonging to the same syntactic category C have content paradigms that are alike, in the sense that there is an invariant set M of morphosyntactic property sets such that for every lexeme L in category C, the content paradigm of L is the complete set of cells pairing L with a member of M. And in the simplest cases, lexemes belonging to the same syntactic category C have form paradigms that are alike, in the sense that for every member  of M there is an 2 Thanks to Olivier Bonami and Grev Corbett for helpful discussion of some of the issues addressed in this chapter.

inflection classes

115

invariant set S of inflectional markings such that for every lexeme L belonging to C, the markings in S are used in the morphological realization of the cell ⟨L, ⟩ in L’s content paradigm. In many languages, however, lexemes belonging to a single syntactic category may fall into different classes, each with its own inventory of inflectional markings. In such instances, lexemes in different classes may have content paradigms that are alike but form paradigms that are different. The classes of lexemes in such instances are inflection classes. Icelandic is a language whose verbs fall into a number of contrasting inflection classes. Verbal lexemes generally have content paradigms in which the lexeme is paired with the thirty morphosyntactic property sets in (1).3 (1)

{ indicative present 1 sg } { indicative present 2 sg } { indicative present 3 sg } { indicative present 1 pl } { indicative present 2 pl } { indicative present 3 pl } { indicative past 1 sg } { indicative past 2 sg } { indicative past 3 sg } { indicative past 1 pl } { indicative past 2 pl } { indicative past 3 pl }

{ subjunctive present 1 sg } { subjunctive present 2 sg } { subjunctive present 3 sg } { subjunctive present 1 pl } { subjunctive present 2 pl } { subjunctive present 3 pl } { subjunctive past 1 sg } { subjunctive past 2 sg } { subjunctive past 3 sg } { subjunctive past 1 pl } { subjunctive past 2 pl } { subjunctive past 3 pl }

{ infinitive } { 1st participle } { 2nd participle } { imperative 2 sg } { imperative 1 pl } { imperative 2 pl }

The lexemes gr´ipa ‘grasp’ and kalla ‘shout’ both have content paradigms of this sort. Their form paradigms, however, are quite different, as Table 6.1 shows. The verb gr´ipa is a ‘strong’ verb; this simply means that it belongs to a particular inflection class, one of whose hallmarks is that it forms its past tense by means of a stem-vowel alternation rather than by means of a dental suffix; this alternation involves the replacement of í /i/ with ei (in the singular past indicative) or with i /i/ (otherwise). The verb kalla ‘shout’, by contrast, is a ‘weak’ verb: it belongs to an inflection class whose members form their past tense by means of a dental suffix -ðV. The dental suffix (which appears in both the indicative and the subjunctive in a weak verb’s paradigm) is not the only morphological marking that distinguishes kalla’s inflection class from that of gr´ipa. For instance, the morphological realization of gr´ipa’s second-person singular past indicative cell involves a suffix -st which is absent from kalla’s form paradigm. In addition, the replacement of í with i is also an exponent of the second participle in gr´ipa’s form paradigm; this marking has no counterpart in kalla’s form paradigm. On the other 3 These are the morphosyntactic property sets of synthetically realized verb forms. Icelandic verbs do, however, have some periphrastic realizations, which I leave aside for present purposes. See Spencer and Popova (this volume) concerning inflectional periphrasis.

116

gregory stump

Table 6.1 Form paradigms of two Icelandic verbs (synthetic forms only)

grípa ‘grasp’

1sg 2sg 3sg 1pl 2pl 3pl

Indicative

Subjunctive

Present Past

Present Past

gríp grípur grípur grípum grípið grípa

greip greipst greip gripum gripuð gripu

grípi grípir grípi grípum grípið grípi

gripi gripir gripi gripum gripuð gripu

kallaði kallaðir kallaði kölluðum kölluðuð kölluðu

kalli kallir kalli köllum kallið kalli

kallaði kallaðir kalla kallaði kölluðum köllum kölluðuð kallið kölluðu

kalla 1sg kalla ‘shout’ 2sg kallar 3sg kallar 1pl köllum 2pl kallið 3pl kalla

Imperative Nonfinite forms gríp

Infinitive: grípa 1st participle: grípandi 2nd participle: gripið

grípum grípið

Infinitive: kalla 1st participle: kallandi 2nd participle: kallað

Source: Jörg (1989).

hand, some inflectional markings are used by both gr´ipa’s inflection class and that of kalla: these include the -r suffix of the second person singular; the first-person plural suffix -um; the -andi suffix of the first participle; and so on.

6.3 Characteristics of Canonical Inf lection Classes ............................................................................................................................................................................. Corbett (2009b) proposes a number of properties as characteristic of canonical4 inflection classes. The value of these properties is that they converge on a typological reference point relative to which actual inflection classes can be compared; actual inflection classes may exhibit a variety of deviations from this canonical ideal. A number of the properties proposed by Corbett are subsumed by a general principle of distinctiveness (‘Canonical inflectional classes are fully comparable and are distinguished as clearly as is possible’, 2009b: 3). Where A and B are inflection classes for lexemes belonging to the same syntactic category in some language, A and B are canonical to the extent that they satisfy the four criteria paraphrased in (2). (2) a. Where RA is the realization of any cell in the content paradigm of a lexeme belonging to class A and RB is the realization of the corresponding cell in 4

For a general discussion of the assumptions and methods of canonical typology, see Corbett (2005).

inflection classes

117

the content paradigm of a lexeme belonging to class B, RA , and RB exhibit a systematic difference in inflectional marking. b. The content paradigms of lexemes belonging to class A and those of lexemes belonging to class B have the same number of cells and their cells are associated with exactly the same morphosyntactic property sets. c. Lexemes belonging to class A are completely uniform in their inflectional realization, as are lexemes belonging to class B; that is, neither class exhibits any internal heterogeneousness in inflectional behaviour. d. All realizations of a lexeme L do an equally good job of identifying L’s inflection-class membership (whether it is a member of class A or of class B).5 In the extreme case in which classes A and B fully satisfy all four criteria, a lexeme LA belonging to class A and a lexeme LB belonging to class B have content paradigms that are completely alike (for each cell ⟨LA , ⟩ in the content paradigm of LA there is a corresponding cell ⟨LB , ⟩ in the content paradigm of LB , and vice versa) but form paradigms that are fully contrastive (the realization of each cell ⟨LA , ⟩ in the content paradigm of LA involves inflectional marking different from that of the realization of the corresponding cell ⟨LB , ⟩ in the content paradigm of LB ).

6.4 Deviations from Canonical Characteristics ............................................................................................................................................................................. Inflection classes satisfying the four criteria in (2) are comparatively rare; deviations from these criterial properties vary widely in kind and complexity. Here I consider both simple and complex deviations.

6.4.1 Simple Deviations A very common deviation involves instances in which lexemes belonging to distinct inflection classes nevertheless exhibit the same morphology in the realization of certain cells in their content paradigms (contrary to (2a)); such cases typically involve morphology that applies, by default, across inflection classes. In Icelandic, for example, the paradigms of gr´ipa and kalla clearly contrast in the first-person singular 5 Criterion (2a) requires a kind of heterogeneousness across inflection classes (the same morphosyntactic property set must differ in its exponence across inflection classes), while criterion (2d) requires a kind of homogeneousness within an inflection class (different morphosyntactic property sets must be alike in the extent to which they are diagnostic of inflection-class membership). In those instances in which both criteria are satisfied, every realization of every lexeme unambiguously identifies that lexeme’s inflection-class membership; in those instances in which (2a) is not satisfied, (2d) entails that all of a lexeme’s realizations are equally unrevealing of its inflection-class membership.

118

gregory stump Table 6.2 Form paradigm of the Icelandic verb SKULU ‘shall’ Indicative Present 1sg 2sg 3sg 1pl 2pl 3pl

Subjunctive Past

skal skalt skal skulum skuluð skulu

Present

Past

skuli skulir skuli skulum skulið skuli

skyldi skyldir skyldi skyldum skylduð skyldu

Imperative

Nonfinite forms Infinitive: 1st participle: 2nd participle:

skulu

Source: Jörg (1989).

Table 6.3 Form paradigm of the Icelandic verb KVÍÐA ‘be anxious’ Indicative

1sg 2sg 3sg 1pl 2pl 3pl

Subjunctive

Present

Past

Present

Past

kvíði kvíðir kvíðir kvíðum kvíðið kvíða

kveið kveiðst kveið kviðum kviðuð kviðu

kvíði kvíðir kvíði kvíðum kvíðið kvíði

kviði kviðir kviði kviðum kviðuð kviðu

Imperative

Nonfinite forms

kvíð

Infinitive: 1st participle: 2nd participle:

kvíða kvíðandi kviðið

kvíðum kvíðið

Source: Jörg (1989).

past indicative (greip vs. kallaði), but in the second-person plural present indicative, the two paradigms are alike: both exhibit the default second-person plural suffix -ið in combination with the default stem (grípið, kallið). Another common deviation involves instances in which lexemes belonging to the same syntactic category have content paradigms that are different, for example in their number of cells (contrary to (2b)). For instance, the content paradigm of the verb skulu ‘shall’ differs from that of most Icelandic verbs in that it lacks imperative, participial, and past indicative cells; see Table 6.2.6 Very often, lexemes whose form paradigms are mostly alike nevertheless differ in minor inflectional details, contrary to criterion (2c). For instance, the verb kv´i-da ‘be anxious’ in Table 6.3 belongs to the same conjugation as gr´ipa ‘grasp’ in Table 6.1 (Conjugation s1a, in the classification of Jörg 1989), yet their form paradigms differ 6

The verb skulu is also unusual in that it is a preterite-present verb; see Section 4.2.4 for discussion.

inflection classes

119

with respect to the endings of the singular present indicative: gríp ‘I grasp’ but kvíði ‘I am anxious’, grípur ‘you (sg) grasp’ but kvíðir ‘you (sg) are anxious’.7 Finally, contrary to criterion (2d), it is not at all unusual for a lexeme’s realizations to differ in the extent to which they reveal its inflection-class membership; for instance, the second-person singular past indicative form greipst in the paradigm of gr´ipa ‘grasp’ in Table 6.1 is a clear mark of its conjugation-class membership, while the second-person plural present indicative form grípið is not.

6.4.2 Complex Deviations Complex deviations from the canonicity criteria in (2) are of various types.

6.4.2.1 Segregated Inflection Classes Some languages exhibit a kind of modularity in their inflection-class systems: in such languages, lexemes belonging to a given syntactic category have content paradigms that are subdivided into sectors, each with its own set of competing inflection classes; these sets are independent in the sense that a lexeme’s membership in a particular inflection class in one set often neither implies nor is implied by its membership in a particular inflection class in another set. In an inflection-class system of this sort, a lexeme simultaneously belongs to more than one inflection class, one for each sector in its paradigm. Finkel and Stump (2007) refer to the inflection classes in a system exhibiting this kind of compartmentalization as segregated inflection classes; I correspondingly refer to the set of segregated inflection classes associated with the same paradigm sector as a segregated set. The ancient Indo-European languages generally have systems of verb inflection in which the conjugations are segregated in this way. In Classical Sanskrit, for example, a verb’s present-system forms (i.e. its present indicative, imperfect, optative, and imperative forms) follow one of the ten conjugations listed in Table 6.4a, but its aorist-system forms (i.e. its aorist indicative forms) follow one of the seven conjugations in Table 6.4b. A verb’s present-system conjugation is for the most part neither predictive of nor predicted by its aorist-system conjugation.8 Table 6.5 presents a typical instance of the sort of cross-classification that arises between present-system and aorist-system conjugations. The verb dhanv ‘run’ is like the verb bhu¯ ‘become’ in the present system, where both inflect as members of the First conjugation. In the aorist, however, they differ: bhu¯ inflects as a member of the Root aorist conjugation, while dhanv inflects as a 7 This difference could, of course, be used to argue that kvi´da and gripa ´ belong to distinct conjugations, in which case their similarities would instead constitute a deviation from (2a). As this case shows, the criteria in (2) relate to inflection-class analyses rather than to unanalysed data. 8 There are some limits on the freedom with which the conjugation classes in Table 6.4 may crossclassify; for instance, the membership of the Tenth conjugation in the present system largely coincides with that of the Reduplicated aorist conjugation in the aorist system (Whitney 1889: §856). This sort of correlation is atypical.

120

gregory stump Table 6.4 Conjugation classes in the Sanskrit present and aorist systems a. Present-system conjugations

b. Aorist-system conjugations

Thematic (stem in -a)

Athematic (stem not in -a)

Asigmatic (stem lacks s suffix)

Sigmatic (stem has s suffix)

First Fourth Sixth Tenth

Second Third Fifth Seventh Eighth Ninth

Root aorist Thematic aorist Reduplicated aorist

s-Aorist is.-Aorist sis.-Aorist sa-Aorist

Table 6.5 Third-person singular indicative active forms of three Sanskrit verbs

Present Aorist

bhu¯ ‘become’

dhanv ‘run’

bhavati abh¯ut

dhanvati adhanv¯ıt

(First) (Root aorist)

ks.an ‘wound’ (First) (is.-Aorist)

ks.an.oti aks.an.¯ıt

(Eighth) (is.-Aorist)

member of the is.-Aorist conjugation. By contrast, dhanv and ks.an ‘wound’ are alike in the aorist (where both inflect as members of the is.-Aorist conjugation) but different in the present, where ks.an inflects as a member of the Eighth conjugation. Segregated sets such as those in Table 6.4 constitute a kind of deviation from the canonical ideal in (2a): since lexemes belonging to distinct inflection classes in one segregated set may belong to the same inflection class in another segregated set, lexemes with partially overlapping inflection-class memberships may inflect differently in the realization of some cells but alike in the realization of others. Segregated sets are also deviations from the canonical ideal in (2d): in Sanskrit, a lexeme’s inflectionclass memberships are not revealed equally clearly by all of its realizations; certain realizations reveal its present-system conjugation-class membership but not that of its aorist-system inflection; others reveal its aorist-system conjugation-class membership but not that of its present-system inflection; and still others (e.g. its past participle) reveal neither its present-system nor its aorist-system conjugation-class membership.

6.4.2.2 Metaconjugation A phenomenon that sometimes arises in systems of segregated conjugations is what Stump (2010) has called metaconjugation. In instances of metaconjugation, the

inflection classes

121

Table 6.6 Imperfect and aorist indicative active conjugations of two Sanskrit verbs

Conjugation

Present system

Imperfect active forms

1 2 3

Indicative active forms

1 2 3

tus. ‘be happy’

Sixth

Fourth

Singular

Dual

Plural

Singular

Dual

Plural

atudam atudas atudat

atud¯ava atudatam atudat¯am

atud¯ama atudata atudan

atus.yam atus.yas atus.yat

atus.y¯ava atus.yatam atus.yat¯am

atus.y¯ama atus.yata atus.yan

Conjugation Aorist system

tud ‘strike’

s-Aorist atautsam atauts¯ıs atauts¯ıt

atautsva atauttam atautt¯am

Sixth atautsma atautta atautsur

atus.am atus.as atus.at

atus.a¯ va atus.atam atus.at¯am

atus.a¯ ma atus.ata atus.an

morphology expressing some lexemes’ membership in a conjugation belonging to one segregated set is also used to express other lexemes’ membership in a conjugation belonging to a different segregated set; that is, the same conjugation appears in distinct segregated sets, serving to realize different sectors of different paradigms. Sanskrit presents a clear example of metaconjugation. In Sanskrit, imperfect forms in the present-system’s Sixth conjugation are morphologically exactly like aorist forms in the Thematic aorist conjugation. Thus, one can equate these as a single conjugation appearing both in the present-system’s segregated set, Table 6.4a, and in the aoristsystem’s segregated set, Table 6.4b. The verb tud ‘strike’ inflects as a member of the Sixth conjugation in the present system and as a member of the s-Aorist conjugation in the aorist system; the verb tus. ‘be happy’, by contrast, inflects as a member of the Sixth conjugation (= Thematic aorist conjugation) in the aorist system and as a member of the Fourth conjugation in the present system. The imperfect and aorist indicative active paradigms in Table 6.6 illustrate. Because the phenomenon of metaconjugation involves segregated sets, it deviates from the canonical ideals in (2a–2d), as observed above. It also deviates from (2c): although tud ‘strike’ and tus. ‘be happy’ both belong to the Sixth/Thematic aorist conjugation, this membership determines the present-system inflection of tud but determines the aorist-system inflection of tus..

6.4.2.3 Heteroclisis Another complex deviation from the properties of canonical inflection classes is heteroclisis. Heteroclisis is similar to the phenomenon of segregated inflection classes in that it involves paradigms divided into sectors such that the inflection of lexeme L in one sector of its paradigm is determined by L’s membership in one inflection class

122

gregory stump Table 6.7 The heteroclite inflection of Czech PRAMEN ‘spring, source’ pokoj (m) ‘room’ Declension

pramen (m) ‘spring’

soft-masculine inanimate

Singular

nom gen dat acc voc loc ins

pokoj pokoje pokoji pokoj pokoji pokoji pokojem

Plural

nom gen dat acc voc loc ins

pokoje pokoj˚u pokoj˚um pokoje pokoje pokojích pokoji

most (m) ‘bridge’

hard-masculine inanimate

pramen pramene prameni pramen prameni prameni pramenem

most mostu mostu most moste mostˇe mostem prameny pramen˚u pramen˚um prameny prameny pramenech prameny

mosty most˚u most˚um mosty mosty mostech mosty

Source: Heim (1982: 22, 41f, 176).

while L’s inflection in another sector of its paradigm is determined by L’s membership in another inflection class. But heteroclisis differs from inflection-class segregation in that the complementary inflection classes in the definition of a heteroclite form paradigm are inflection classes that ordinarily encompass entire paradigms.9 Consider, for example, the Czech noun pramen ‘spring, source’. In the singular, pramen exhibits the soft-masculine inanimate declension of a noun such as pokoj ‘room’; in the plural, it instead exhibits the hard-masculine inanimate declension of a noun such as most ‘bridge’. The paradigms in Table 6.7 illustrate. Heteroclisis is a deviation from the canonical property in (2a): the morphological realizations in pramen’s form paradigm are neither always like nor always different from those in the form paradigms of pokoj and most. Heteroclisis is by definition a deviation from (2c): in the example at hand, Czech nouns belonging to the soft-masculine inanimate declension are not completely uniform in their inflectional behaviour, since at least one of these is heteroclite. Heteroclitic paradigms also deviate from (2d): even if pramen’s singular forms reveal its membership in the soft-masculine inanimate declension, they do not reveal whether its plural forms will follow the pattern of pokoj or that of most. 9

See Stump (2006) for extensive discussion of the properties of heteroclisis.

inflection classes

123

6.4.2.4 Heteroclisis with Deponency In an especially complex kind of deviation from the properties of canonical inflection classes, heteroclisis coincides with deponency. In the familiar instance of Latin deponency, a deponent verb’s active forms exhibit the morphology reserved for passive forms in the form paradigms of ordinary verbs: for example deponent hortor ‘I urge’, cf. nondeponent laudor ‘I am being praised’. Thus, the term ‘deponency’ (deponere ‘lay aside’) describes inflectional morphology that has laid aside its usual function (in the Latin case, the expression of passive voice) in favour of another (the expression of active voice).10 In Icelandic, certain special verbs inflect in the present tense by means of morphology that other verbs use in the past tense. This is especially clear in Old Icelandic, where these special verbs—the preterite-present verbs (eiga ‘have’, knattu ´ ‘know, be able’, kunna ‘know, be able’, mega ‘be able’, muna ‘remember’, munu ‘will’, skylu ‘shall’, þurfa ‘need’, unna ‘love’, vita ‘know’)—employ the morphology that is usual for a strong verb’s past-tense forms; the Old Icelandic paradigms in Table 6.8 illustrate. Icelandic preterite-present verbs are analogous to Latin deponents: just as the latter inflect in the active voice by means of morphology whose usual passive function is laid aside, the former inflect in the present tense by means of morphology whose usual past-tense function is laid aside. There is, however, a difference: in Latin, deponent verbs do not have finite passive forms, but in Icelandic, preterite-present verbs generally do have past-tense forms. These are formed by means of morphology that is elsewhere used in the past-tense inflection of weak verbs. Thus, Icelandic preterite-present verbs are both deponent and heteroclite: they form their present tense as strong verbs form their past tense, and they form their past tense as weak verbs form their past tense. The reader can verify that this inflection deviates from the canonical ideals in (2a), (2c), and (2d).

6.5 Expressions of Inf lection-class Membership ............................................................................................................................................................................. Logically, there are two sorts of extremes for the morphological expression of inflection-class membership. At one extreme, a lexeme’s inflection-class membership is expressed entirely by the inventory of exponents with which its stem is marked; at this extreme, lexemes belonging to different inflection classes contrast not in the kinds of stems they possess, but in the inventories of exponents with which those stems are marked. At the opposite extreme, contrasting inflection classes are associated with a common set of inflectional exponents but contrast by virtue of their distinct patterns of stem alternation. It is unusual for a language’s inflection classes to be purely of one 10

For a general discussion of systems of deponency, see Baerman (2007b, this volume).

Table 6.8 Indicative and subjunctive paradigms of three Old Icelandic verbs (Shaded forms of þURF Ainflect like shaded forms of BRENNA; heavy-bordered boxes enclose forms of þURFA and DUGA that inflect alike.) Strong BRENNA‘burn’

Indicative

Pres.

SG PL

Past

SG PL

Subjunctive Pres.

SG PL

Past

SG PL

Source: Zoëga (1910).

brenn, brenn-r, brenn-r brenn-um, brenn-ið, brenn-a brann, brann-t, brann brunn-um, brunn-uð, brunn-u brenn-a, brenn-ir, brenn-i brenn-im, brenn-ið, brenn-i brynn-a, brynn-ir, brynn-i brynn-im, brynn-ið, brynn-i

Preterite-present ÞURFA‘need’ þarf, þarf-t, þarf þurf-um, þurf-uð, þurf-u þurf-ta, þurf-tir, þurf-ti þurf-tum, þurf-tuð, þurf-tu þurf-a, þurf-ir, þurf-i þurf-im, þurf-ið, þurf-i þyrf-ta, þyrf-tir, þyrf-ti þyrf-tim, þyrf-tið, þyrf-ti

Weak DUGA‘help’

dug-i, dug-ir, dug-ir dug-um, dug-ið, dug-a dug-ða, dug-ðir, dug-ði dug-ðum, dug-ðuð, dug-ðu dug-a, dug-ir, dug-i dug-im, dug-ið, dug-i dyg-ða, dyg-ðir, dyg-ði dyg-ðim, dyg-ðið, dyg-ði

inflection classes

125

Table 6.9 Form paradigms of two strong verbs in Icelandic Indicative Present grípa ‘grasp’

Subjunctive Past

Present

Past

Imperative

Nonfinite forms

1sg gríp 2sg gríp-ur 3sg gríp-ur

greip greip-st greip

gríp-i gríp-ir gríp-i

grip-i grip-ir grip-i

gríp

Infinitive: gríp-a 1st participle: gríp-andi 2nd participle: grip-ið

1pl gríp-um 2pl gríp-ið 3pl gríp-a

grip-um gríp-um grip-uð gríp-ið gripu gríp-i

grip-um gríp-um grip-uð gríp-ið grip-u

braut brau-st braut

bryt-i bryt-ir bryt-i

brjóta 1sg brýt ‘break’ 2sg brýt-ur 3sg brýt-ur

brjót-i brjót-ir brjót-i

brjót

Infinitive: brjót-a 1st participle: brjót-andi 2nd participle: brot-ið

1pl brjót-um brut-um brjót-um bryt-um brjót-um 2pl brjót-ið brut-uð brjót-ið bryt-uð brjót-ið 3pl brjót-a brut-u brjót-i bryt-u Source: Jörg (1989).

or the other of these two extreme types; much more commonly, a language’s inflection classes blend the characteristics of the extremes. Icelandic is a case in point. The class of Icelandic strong verbs subsumes a number of subclasses each of which constitutes a separate conjugation. The differences among these conjugations are generally different patterns of stem alternation. A comparison of the finite conjugation of gr´ipa ‘grasp’ with that of brjóta ‘break’ (Table 6.9) reveals that these verbs employ the same inventory of inflectional endings; their difference is a difference of stem alternation. The conjugation of gr´ipa employs the ablaut pattern in (3a), while the conjugation of brjóta ‘break’ instead exhibits the pattern in (3b). (3)

Singular Singular Plural Subjunctive General 2nd present — past — past — — — past default participle indicative indicative indicative a. í ei i i í i b. ý au u y jó o

Weak verbs also fall into a number of subconjugations; these are distinguished less by patterns of stem alternation than by their inventories of endings. Consider, for example, the form paradigms of the verbs kalla ‘shout’ and fryja ´ ‘challenge’ (Table 6.10): while kalla employs the ending -a in the first singular present indicative and in the second singular imperative, fryja ´ is endingless in both forms; while kalla employs the ending -ar in the second singular present indicative, fryja ´ instead employs the

126

gregory stump

Table 6.10 Form paradigms of two weak verbs in Icelandic

kalla ‘shout’

Indicative

Subjunctive

Present Past

Present Past

Imperative Nonfinite forms

kall-i kall-ir kall-i

kall-a

1sg kall-a kall-aði 2sg kall-ar kall-aðir 3sg kall-ar kall-aði

kall-aði kall-aðir kall-aði

Infinitive: kall-a 1st participle: kall-andi 2nd participle: kall-að

1pl köll-um köll-uðum köll-um köll-uðum köll-um 2pl kall-ið köll-uðuð kall-ið köll-uðuð kall-ið 3pl kall-a köll-uðu kall-i köll-uðu fryja ´ 1sg frý ‘challenge’ 2sg frý-rð 3sg frý-r

frý-ði frý-ðir frý-ði

1pl frý-jum frý-ðum 2pl frý-ið frý-ðuð 3pl frý-ja frý-ðu

frý-i frý-ir frý-i

frý-ði frý-ðir frý-ði

frý-jum frý-ðum frý-ið frý-ðuð frý-i frý-ðu

frý

Infinitive: frý-ja 1st participle: frý-jandi 2nd participle: frý-ð, frý-jað

frý-jum frý-ið

Source: Jörg (1989).

ending -rð; and so on. The stems of these verbs remain largely constant; neither exhibits a pattern of ablaut comparable to those in (3). (The verb kalla does, however, exhibit u-mutation: the stem-internal vowel a becomes ö before a suffixal u.) Thus, Icelandic conjugations vary in the extent to which they are distinguished by special patterns of stem marking and special patterns of stem alternation.

6.6 Correlates of Inf lection-class Membership ............................................................................................................................................................................. Besides embodying the principle of distinctiveness discussed in Section 6.3, canonical inflection classes also embody a second general principle under Corbett’s assumptions. This is the principle of independence, according to which ‘[t]he distribution of lexical items over canonical inflectional classes is synchronically unmotivated’ (Corbett 2009b: 5). By this principle, a lexeme’s inflection-class membership is not, in the canonical case, a function of its phonology, syntax, semantics, or pragmatics. Many inflection-class systems exhibit this kind of independence to some degree. In Icelandic, for example, the phonology of a verbal lexeme’s stem is not, in general, a predictor of its inflection-class membership; thus, aka ‘drive’ and skaka ‘shake’ belong to strong

inflection classes

127

Table 6.11 General forms of three verbs in Moru (Miza dialect) Class I zi ‘call’

Class II εmε ‘heat’

Class III kanda ‘shake’

1sg 2sg 3sg

mä´-zi mí-zi ányä zi

m-´ε mε ny-´εmε ány-εmε

má-kanda nyá-kanda ánya kanda

1pl 2pl 3pl

mà-zi mì-zi ànyä zi

m-´εmε ny-´εmε àny-εmε

mà-kanda nyà-kanda ànya kanda

Where v is a vowel symbol, v´ indicates high tone, v` low tone, and v mid tone; ä represents a high central vowel. Source: Tucker (1940).

conjugation 6b,11 while baka ‘bake’ and raka ‘shave’ belong to weak conjugation 4a, and flaka ‘be loose’ and vaka ‘be awake’ to weak conjugation 3d. Nevertheless, correlations between a lexeme’s inflection-class membership and its other properties are by no means unusual. Consider a pair of examples. In Moru (Nilo-Saharan; Sudan), verbs fall into three conjugations whose membership is phonologically determined: monosyllabic verbs belong to Class I, vowel-initial disyllabic verbs to Class II, and consonant-initial disyllabic verbs belong to Class III. Examples are the verbs in Table 6.11; as these examples suggest, Classes II and III can be identified as a single class by assuming that a prefixal vowel is elided prevocalically. In Muna (Austronesian; Indonesia), verbs fall into three conjugations, which van den Berg (1989) calls the a-Class, the ae-Class, and the ao-Class according to the form of the first-person singular subject prefix in each conjugation. Contrasts such as a-losa ‘I emerge’ vs. ae-lobhi ‘I hit’ vs. ao-lowu ‘I am drunk’ show that membership in these conjugations is not phonologically determined. Instead, conjugation-class membership correlates with the syntactico-semantic properties of transitivity and stativity: a-Class verbs tend to be dynamic intransitives, ae-Class verbs tend to be transitives, and ao-Class verbs tend to be stative intransitives. But these correlations are not perfect, as the figures in Table 6.12 show; rather, they must be seen as default correspondences subject to lexical override. Even if a lexeme’s inflection-class membership is not predictable from its other properties, its inflection-class membership may be a predictor of other properties. The well-known correlations between declensions and gender in Latin have already been alluded to. In Sanskrit, verbs fall into three groups: ubhayapadin verbs have contrasting active-voice and middle-voice inflections; parasmaipadin verbs have active-voice inflections only; and a¯ tmanepadin verbs have middle-voice inflections only. In the 11

The conjugation-class labels are those of Jörg (1989).

128

gregory stump Table 6.12 Syntax/semantics of verbs in different Muna conjugations Conjugation

Syntax/semantics

Verbs

Proportion of all verbs

a-Class

transitive dynamic intransitive stative intransitive

5 46 15

2.5% ⌉ 23.0% ⊢ (= 33%) 7.5% ⌋

ae-Class

transitive dynamic intransitive stative intransitive

80 13 3

40.0% ⌉ 6.5% ⊢(= 48%) 1.5% ⌋

ao-Class

transitive dynamic intransitive stative intransitive

1 5 32

0.5% ⌉ 2.5% ⊢ (= 19%) 16.0% ⌋

Source: van den Berg (1989: 55f).

Table 6.13 Nine Sanskrit verbs and their aorist conjugations

Root aorist Thematic aorist Reduplicated aorist s-aorist is.-aorist sis.-aorist sa-aorist

Active

Middle

d¯a ‘give’, bhu¯ ‘become’ chid ‘cut’, vr.dh ‘grow’, sic ‘pour’ mocaya ‘cause to release’ chid ‘cut’, n¯i ‘lead’ pu¯ ‘purify’ jña ‘know’ duh ‘milk’

(none) sic (few such lexemes over all) mocaya chid, jña, d¯a, n¯ı ¯ bhu, ¯ vr.dh pu, (none) duh

aorist, certain conjugations are restricted (wholly or nearly) to the active voice, while none is restricted to the middle voice (Whitney 1889: §§829, 846, 911). In particular, verbs that form their active voice as members of the Root aorist conjugation form their middle voice as members of the s- or is.-Aorist conjugation (the same is largely true of verbs forming their active voice as members of the Thematic aorist conjugation), and verbs forming their active voice as members of the sis.-Aorist conjugation form their middle as members of the s-Aorist conjugation. The examples in Table 6.13 illustrate. Thus, verbs that follow the Root aorist or sis.-Aorist conjugation cannot be a¯ tmanepadin, and the aorist paradigms of ubhayapadin verbs that follow the Root aorist or sis.-Aorist conjugation must be heteroclite. Because rules of word derivation often determine both the meaning and the inflection-class membership of derived words, a language’s derivational morphology may establish correlations between a derived lexeme’s inflection-class membership and

inflection classes

129

Table 6.14 Some Sanskrit verbs and their causative derivatives in -aya Sample lexeme

vad dvis. d¯a sampad s´ ru ks.ip yuj kr. grah

Simple present-system stem

‘speak’ ‘hate’ ‘give’ ‘succeed’ ‘hear’ ‘throw’ ‘yoke’ ‘make’ ‘seize’

Form

Conjugation

vadadves.dad¯asampadya´sr.n.oks.ipayunajkarogr.hn.a¯ -

First Second Third Fourth Fifth Sixth Seventh Eighth Ninth

Causative stem (Tenth conjugation) v¯adayadves.ayad¯apayasamp¯adaya´sr¯avayaks.epayayojayak¯arayagr¯ahaya-

its semantic properties. In Sanskrit, for example, a verb’s causative derivative is formed by a shift to the Tenth conjugation; consequently, Tenth-conjugation verbs tend to have causative meanings. The examples in Table 6.14 illustrate; in this table, alternating stems (e.g. karo- ~ kuru-, gr.hn.a¯ - ~ gr.hn.¯ı-) are given in their strong form. When most of the lexemes in an inflection class belong to a coherent class of derivatives, the marks of their inflection-class membership are sometimes mistakenly regarded as marks of the associated derivational category. Thus, the suffix -aya in Table 6.14 is sometimes portrayed as a causative derivational suffix. Three observations are relevant to evaluating this portrayal. First, Tenth-conjugation verbs with stems in -aya are not always derivatives; for example, cur ‘steal’ (present-tense stem coraya-) is a basic member of the Tenth conjugation. Second, derived Tenth-conjugation verbs with stems in -aya do not always have a causative meaning; for example, the verbs kr.p ‘lament’ and il ‘be quiet’ (both members of the Sixth conjugation, with present-tense stems kr.pa- and ila-) both have Tenth-conjugation derivatives (stems kr.paya- ‘lament’, ilaya- ‘cease’), neither of which has a causative meaning. The problem posed by cur might be reconciled with the assumption that -aya is a causative derivational suffix by the observation that classes of derived forms often include occasional forms for which there is no synchronic base (e.g. furious has fury as its base, but curious lacks any synchronic base), and the problem posed by kr.p and il might be reconciled by the observation that classes of derived forms often include occasional forms whose semantics is idiosyncratic (as comfortable is idiosyncratic alongside breakable). But a third property of -aya is less easily countered: while a derived lexeme’s derivational markings are ordinarily preserved throughout that lexeme’s paradigm, -aya only appears in part of a derived causative’s paradigm. Thus, derived desiderative verbs in Sanskrit are marked by a reduplicative prefix and a sibilant suffix, and these are preserved throughout their paradigm; but a causative verb’s stem in -aya is restricted to certain cells of

130

gregory stump

Table 6.15 Finite (third person singular) and nonfinite forms of the desiderative derivative of KR. ‘make, do’

Present system:

Present Imperfect Imperative Optative

General tenses:

Aorist Perfect -sya Future Periph. Future Conditional Benedictive

Present participial stem:

active middle

Past passive participial stem Gerund: Gerundive stem: Infinitive

in -tv¯a in -ya in -ya in -tavya

Active

Middle

Passive

cik¯ırs.ati acik¯ırs.at cik¯ırs.atu cik¯ırs.et acik¯ırs.¯ıt cik¯ırs.a¯ m a¯ sa cik¯ırs.is.yati cik¯ırs.it¯a acik¯ırs.is.yat cik¯ırs.y¯at

cik¯ırs.ate cik¯ırs.yate acik¯ırs.ata acik¯ırs.yata cik¯ırs.at¯am cik¯ırs.yat¯am cik¯ırs.eta cik¯ırs.yeta acik¯ırs.is..ta acik¯ırs.i cik¯ırs.a¯ m a¯ se cik¯ırs.is.yate cik¯ırs.it¯a acik¯ırs.is.yata cik¯ırs.is.¯ı.s.ta

cik¯ırs.antcik¯ırs.am¯anacik¯ırs.itacik¯ırs.itv¯a -cik¯ırs.ya cik¯ırs.yacik¯ırs.itavyacik¯ırs.itum

All forms exhibit identical desiderative morphology (reduplicative prefix ci- and sibilant suffix).

its paradigm. This difference is illustrated in Table 6.15 with cik´irs. ‘want to make/do’ (the desiderative derivative of kr. ‘make/do’) and in Table 6.16 with k¯araya ‘cause to make/do’ (the causative derivative of kr. ‘make/do’). The inevitable conclusion is that -aya is a mark of inflection-class membership rather than a derivational suffix (Stump 2005a).

6.7 The Evolution of Inf lection Classes

.............................................................................................................................................................................

Inflection classes present three interesting problems in the diachronic dimension of linguistic analysis. First, how do inflection-class distinctions arise? Second, how can a lexeme change its inflection-class membership? Third, how are inflection-class distinctions lost? Despite the naïve intuition that inflection-class distinctions do not serve any purpose and would be dispensed with in an intelligently designed language, inflection classes are extremely hardy; inflection-class distinctions may persist for millennia. New inflection-class distinctions arise in various ways.

inflection classes

131

Table 6.16 Finite (third person singular) and nonfinite forms of the causative derivative of KR. ‘make, do’ Active

Middle

Passive

k¯arayate ak¯arayata k¯arayat¯am k¯arayeta

k¯aryate* ak¯aryata* k¯aryat¯am* k¯aryeta*

Present system:

Present Imperfect Imperative Optative

k¯arayati ak¯arayat k¯arayatu k¯arayet

General tenses:

Aorist Perfect -sya Future Periph. Future Conditional Benedictive

ac¯ıkarat* k¯aray¯añ cak¯ara k¯arayis.yati k¯arayit¯a ak¯arayis.yat k¯ary¯at*

Present participial stem:

active middle

k¯arayantk¯arayam¯ana-

Past passive participial stem

k¯arita-*

Gerund:

in -tv¯a in -ya

k¯arayitv¯a -k¯arya*

Gerundive stem:

in -ya in -tavya

k¯arya-* k¯arayitavya-

Infinitive

ac¯ıkarata* k¯aray¯añ cakre k¯arayis.yate k¯arayit¯a ak¯arayis.yata k¯arayis.¯ı.s.ta

k¯arayitum

*Asterisked forms do not exhibit the suffix -aya (or one of its sandhi variants).

Sound changes may obscure the similarity of forms that inflect alike to such an extent that they are reanalysed as belonging to distinct inflection classes. In Vedic, the verbs yuj ‘harness’ and pu¯ ‘purify’ belong to two distinct conjugations: yuj belongs to the Seventh conjugation, while pu¯ belongs to the Ninth conjugation; their present indicative active forms are given in Table 6.17. Verbs belonging to the Seventh conjugation form their present-system stem by means of a nasal infix: -na- in strong forms (e.g. the singular of the present indicative active), -n- in weak forms (e.g. nonsingular forms of the present indicative active). Thus, the present-system stem of yuj is yu-na-j- (strong; sandhi form yunak-) or yu-n-j- (weak; sandhi forms yuñj-, yu˙nk-). Verbs belonging to the Ninth conjugation form their present-system stem by means of the suffix -n¯a (strong forms) or -n¯ı (weak forms); thus, the present-system stem of pu¯ is pun¯a- (strong) or pun¯ı- (weak; the ¯ı is elided prevocalically in pun-anti ‘they purify’). The Vedic roots yuj and p¯u are presumed to descend from the Proto-Indo-European roots *iug ̯ and *puH, where h is a laryngeal (Mayrhofer 1972: §106; Lindeman 1987: §13). According to the Saussurean analysis, these belonged to the same conjugation, forming their present-system stems through the infixation of *-ne- (strong forms) or *-n- (weak forms): thus, *iu-ne-g-, ̯ *pu-ne-H- (strong) and *iu-n-g-, ̯ *pu-n-H- (weak).

132

gregory stump

Table 6.17 Present indicative active forms of two Vedic verbs

1 2 3

yuj ‘harness’, Seventh conjugation

pu¯ ‘purify’, Ninth conjugation

Singular

Dual

Plural

Singular

Dual

Plural

yunaj-mi yunak-s.i yunak-ti

yuñj-vas yu˙nk-thas yu˙nk-tas

yuñj-mas yu˙nk-tha yuñj-anti

pun¯a-mi pun¯a-si pun¯a-ti

pun¯ı-vas pun¯ı-thas pun¯ı-tas

pun¯ı-mas pun¯ı-tha pun-anti

This conjugation becomes the Seventh conjugation in Vedic. But sound changes caused *-ne-H to become Vedic -n¯a and caused *-n-H to become Vedic -n¯ı (preconsonantally) or -n (prevocalically); as a consequence, the connection of the stems pun¯a-/pun¯ı- (and others like them) to the stem-formation pattern of the Seventh conjugation became opaque. The ancient Sanskrit grammarians treated them as embodying a distinct, Ninth conjugation, suggesting that speakers had reanalysed stems like pun¯a-/pun¯ı- as involving a suffixal pattern rather than the inherited infixal pattern of the Seventh conjugation. Thus, in this case, sound changes apparently led to the split of an inherited inflection class into two distinct classes. New inflection-class distinctions may also arise through the reanalysis of elements originating in other functions. Indo-European is assumed to have had a suffix *-ske/o for the derivation of verbs with an iterative or durative meaning (Szemerényi 1996: 273). In some Indo-European languages, the reflex of *-ske/o is a suffix bleached of any specifically iterative/durative significance but serving as the stem formative of a particular inflection class. For instance, it surfaces in Vedic as -ccha, which forms the present-system stems of a small subclass of First-conjugation verbs (Mayrhofer 1972: §§18, 83): gam ‘go’, gacchati ‘s/he goes’; yam ‘restrain’, yacchati ‘s/he restrains’; yu ‘separate’, yucchati ‘s/he separates’. It surfaces in Italian as the suffix -sc, which appears in the singular and third plural present-tense forms of a subclass of Thirdconjugation verbs, for example capire ‘understand’: capisco ‘I understand’, capiscono ‘they understand’, but capiamo ‘we understand’. Analogical processes naturally produce shifts in inflection-class membership; for instance, Middle-English speakers extended the ablaut pattern of bear—bore, swear— swore (strong verbs) to the weak verb wear, whose past tense wered was replaced with wore. Changes of this sort need not be abrupt; they may spread gradually in a lexeme’s paradigm, through intermediate stages involving heteroclite inflection. For example, in Old Indic (here exemplified by Sanskrit), neuter nouns with stems in as, such as manas ‘mind’, follow the distinctive declensional pattern in the leftmost column of Table 6.18; in Middle Indic (here exemplified by P¯ali), these nouns give evidence of ¯ shifting to the neuter a-stem declension, exemplified in Table 6.18 by rupa ‘material object’ (Geiger 1994: §99; Perniola 1997: §196). Thus, in the P¯ali paradigm of manas in Table 6.18, the shaded forms (i.e. the plurals and the ablative singular

inflection classes

133

form man¯a) do not follow the expected declensional pattern reflected in Sanskrit; instead, they follow the pattern of neuter a-stem nouns, as if their stem were manarather than manas-. These facts suggest that a gradual transfer of manas and nouns like it to the a-stem declension is in progress in P¯ali (a hypothesis confirmed by the occasional incidence of additional singular a-stem forms in the paradigms of P¯ali asstems). This change must have been motivated by the extremely high incidence of a-stem nouns in P¯ali. It may also have been spurred by the fact that, notwithstanding the difference of gender, the traditional inflection of as-stem nouns coincides with that of masculine a-stem nouns (e.g. deva ‘god’) in the nominative singular: thus, nom sg devo ‘god’ is to the inherited form mano as gen pl dev¯anam . is to the 12 innovative form man¯anam . ; as Table 6.18 shows, the declension of deva in P¯ali provides an analogical basis for all of the innovative oblique forms (plural and singular) of manas. Processes of language change can completely eliminate inflection-class distinctions. For instance, feminine i-stems, monosyllabic feminine ¯ı-stems, and polysyllabic feminine ¯ı-stems belong to three distinct declensions in Sanskrit, as the paradigms in Table 6.19 show. In P¯ali, these three declensions merge (Bubenik 1996: 79f; Perniola 1997: §195; Stump 2001a). In the resulting declension, lexical specification determines the quantity of the stem-final vowel in the nominative/vocative singular; outside the nominative/vocative singular, stem-final ¯˘ı shows the pattern of phonologically conditioned alternation observed in the Sanskrit inflection of dh¯i (iy before vowels, ¯ı elsewhere); and the endings are the reflexes of those observed with Sanskrit nad¯i (whose influence on the declensions of j¯ati and dh¯i is already observable in variant forms in Sanskrit). Evidence of this sort poses the familiar problem of actuation in historical linguistics: why did the inflection-class distinction between the j¯ati, dh¯i, and nad¯i declensions, which had persisted for centuries in Old Indic, disintegrate in the transition to Middle Indic? An explanatory account of this development must likely involve the complex interaction of a number of contributory factors, including sound change, language contact and the high frequency of polysyllabic ¯ı-stems.

6.8 Inf lection Classes in Linguistic Theory and Typology ............................................................................................................................................................................. In formal models of natural language, inflection classes pose a problem of representation. Given a lexeme L in the lexicon of some language, how is the inflection-class membership of L to be expressed? One option is to represent L’s inflection-class membership by means of a label such as [Strong verb] or [Second declension]; on this 12 The P¯ ali paradigms in Table 6.18 exhibit two innovative patterns of syncretism: that of the dative with the genitive and that of the ablative with the instrumental.

134

gregory stump

¯ Table 6.18 Declension of MANAS ‘mind’ (neut.), R UPA ‘material object’ (neut.) and DEVA ‘god’ (masc.) in Sanskrit and P¯ali

Sanskrit

P¯ali

Singular nom voc acc ins dat abl gen loc

manas manas manas manas¯a manase manasas manasas manasi

r¯upam r¯upam r¯upam r¯upena r¯up¯aya r¯up¯at r¯upasya r¯upe

devas deva devam devena dev¯aya dev¯at devasya deve

Plural

man¯am . si man¯am . si man¯am . si manobhis manobhyas manobhyas manas¯am manah.su

r¯up¯ani r¯up¯ani r¯up¯ani r¯upais* r¯upebhyas r¯upebhyas r¯up¯an¯am r¯upes.u

dev¯as dev¯as dev¯an devais* devebhyas devebhyas dev¯an¯am deves.u

nom voc acc ins dat abl gen loc

mano mano mano manas¯a manaso manas¯a, man¯a manaso manasi man¯ani man¯ani man¯ani manehi man¯anam . manehi man¯anam . manesu

r¯upam . r¯upam . r¯upam . r¯upena r¯upassa r¯up¯a r¯upassa r¯upe

devo deva devam . devena devassa dev¯a devassa deve

r¯up¯ani r¯up¯ani r¯up¯ani r¯upehi r¯up¯anam . r¯upehi r¯up¯anam . r¯upesu

dev¯a dev¯a deve devehi dev¯anam . devehi dev¯anam . devesu

* Vedic also r¯upebhis, devebhis.

option, the rules of morphology defining L’s form paradigm are conditioned by reference to this label. A different option is to specify a set of principal parts for L: key members of L’s form paradigm that allow the other members to be deduced by language-specific patterns of inference.

6.8.1 Inflection-Class Labels The label solution to the problem of representing inflection-class membership has the advantage of economy: a single label such as [Fourth declension] is shorter and simpler than a sequence of principal parts such as fr¯uctus—fr¯uct¯us. The postulation of inflection-class labels does, however, require a choice: should the label be seen as the name of a set (as in Stump 2001b) or is its theoretical status that of a diacritic feature comparable to the specification of a morphosyntactic property (as in Corbett and Baerman 2006)?13 A possible argument for assimilating inflection-class labels 13

In Network Morphology (Corbett and Fraser 1993; Brown and Hippisley 2012), a language’s inflectional system is modelled as a default inheritance hierarchy in which inflection classes are sharply distinguished from morphosyntactic property specifications: while the latter are atoms used in the formulation of inheritable morphological rules, inflection classes are formalized as nodes defining a hierarchy’s patterns of inheritance. In this approach, membership in a particular inflection class is represented as default inheritance from a particular node.

inflection classes

135

Table 6.19 Declension of feminine i- and ¯ı -stem nouns in Sanskrit and P¯ali i-stem: j¯ati ‘birth’

monosyllabic ¯ı-stem: dh-i ‘thought’

polysyllabic ¯ı-stem: nad-i ‘river’

Singular

Plural

Singular

Plural

Singular

Plural

Sanskrit

nom voc acc ins dat abl gen loc

j¯atis j¯ate j¯atim j¯aty¯a j¯ataye* j¯ates* j¯ates* j¯atau*

j¯atayas j¯atayas j¯at¯ıs j¯atibhis j¯atibhyas j¯atibhyas j¯at¯ın¯am j¯atis.u

dh¯ıs dh¯ıs dhiyam dhiy¯a dhiye** dhiyas** dhiyas** dhiyi**

dhiyas dhiyas dhiyas dh¯ıbhis dh¯ıbhyas dh¯ıbhyas dhiy¯am** dh¯ı.su

nad¯ı nadi nad¯ım nady¯a nadyai nady¯as nady¯as nady¯am

nadyas nadyas nad¯ıs nad¯ıbhis nad¯ıbhyas nad¯ıbhyas nad¯ın¯am nad¯ı.su

P¯ali

nom voc acc ins dat abl gen loc

j¯ati j¯ati j¯at¯ım . j¯atiy¯a j¯atiy¯a j¯atiy¯a j¯atiy¯a j¯atiyam ., j¯atiy¯a

j¯atiyo, j¯at¯ı j¯atiyo, j¯at¯ı j¯atiyo, j¯at¯ı j¯at¯ıhi j¯at¯ınam . j¯at¯ıhi j¯at¯ınam . j¯at¯ısu

dh¯ı dh¯ı dh¯ım . dhiy¯a dhiy¯a dhiy¯a dhiy¯a dhiyam ., dhiy¯a

dhiyo, dh¯ı dhiyo, dh¯ı dhiyo, dh¯ı dh¯ıhi dh¯ınam . dh¯ıhi dh¯ınam . dh¯ısu

nad¯ı nad¯ı nad¯ım . nadiy¯a nadiy¯a nadiy¯a nadiy¯a nadiyam ., nadiy¯a

nadiyo, nad¯ı nadiyo, nad¯ı nadiyo, nad¯ı nad¯ıhi nad¯ınam . nad¯ıhi nad¯ınam . nad¯ısu

Variant inflections similar to that of nad-i: *dat sg j¯atyai, abl/gen sg j¯aty¯as, loc sg j¯aty¯am; **dat sg dhiyai, abl/gen sg dhiy¯as, loc sg dhiy¯am, gen pl dh¯ın¯am.

to morphosyntactic property specifications is that they behave alike in resolving rule competition. Consider an example. In the inflection of Icelandic verbs, the rules in (4) are among those that realize second-person singular subject agreement. Rule (4c) is overridden by rule (4b), which is in turn overridden by (4a). This pattern of overrides can be attributed to principle (5) on the assumption that rules (4a–4c) realize the respective property sets in (6). On this analysis, the inflection-class label X has the same role in resolving rule competition as the morphosyntactic property specifications ‘2’, ‘sg’, and ‘imperative’. (4) a. Verbs in class X are suffixed with -ðu in the second-person singular imperative. b. Verbs are suffixless in the second-person singular imperative. c. Verbs are suffixed with -r in the second-person singular. (5) A rule realizing a set S of properties overrides any competing rule realizing a proper subset of S.

136

gregory stump

Table 6.20 Present indicative paradigms of four Romanian verbs Conjugation: Exemplar:

First a invita ‘to invite’

Second ˘ a tacea ‘to be silent’

Third a face ‘to do’

Fourth a s¸ ti ‘to know’

1sg 2sg 3sg

invít inví¸ti invít˘a

tác táci táce

fác fáci fáce

¸stíu ¸stíi ¸stíe

1pl 2pl 3pl

invít˘am invít˘a¸ti invít˘a

t˘acém t˘acé¸ti tác

fácem fáce¸ti fác

¸stím ¸stí¸ti ¸stíu

Source: Stump (2001b: 214).

(6) a. {X, 2, sg, imperative} b. {2, sg, imperative} c. {2, sg} It is not certain, however, that inflection-class labels actually have the same status as morphosyntactic property specifications in the resolution of rule competition. Stump (2001b: 52) appeals to the principle in (7) to mediate instances of rule competition. The notion of narrowness invoked by (7) is defined relative to the inflection class to whose members a rule R applies (conceived of as a set SR ) and to the morphosyntactic property set R that R realizes through its application: (a) if SR1 = SR2 , then rule R1 is narrower than rule R2 if and only if R2 is a proper subset of R1 ; (b) if SR1 ̸= SR2 , then R1 is narrower than R2 if and only if SR1 is a subset of SR2 . This conception of narrowness entails that inflection-class labels have a different status from morphosyntactic property specifications: in particular, it entails that if R1 is a proper subset of R2 but SR1 is a subset of SR2 , then the application of R1 overrides that of R2; that is, subset relations among inflection classes weigh more heavily than subset relations among morphosyntactic property sets in the resolution of rule competition. (7) In instances of rule competition, the narrower rule overrides its competitor. Romanian provides evidence that apparently confirms this contrast between inflectionclass labels and morphosyntactic property specifications. In Romanian, the default rule of third-person singular agreement marking is (8a); this rule applies in verbs in the Second, Third, and Fourth conjugations, but is overridden in the First conjugation by rule (8b). (The forms in Table 6.20 illustrate.) The morphosyntactic property set {3} realized by rule (8b) is a proper subset of the set {3, sg} realized by (8a), but (8b) nevertheless overrides (8a) because the set of First-conjugation verbs is a proper subset of verbs. This is prima facie evidence that the grammatical status of inflection-class labels should not be equated with that of morphosyntactic property specifications.

inflection classes

137

(8) a. Verbs associated with the morphosyntactic property set {3, sg} are suffixed with -e. b. First-conjugation verbs associated with the morphosyntactic property set {3} are suffixed with -˘a.

6.8.2 Principal Parts and Inflection Classes A problem with the label solution to the problem of representing inflection-class membership is that labels fail to embody language users’ awareness that the words in a lexeme’s form paradigm are bound together by a network of implicative relations (Wurzel 1989: 112ff; Blevins 2006; Finkel and Stump 2007). These relations allow language users to deduce a lexeme’s full paradigm of inflected forms from a few key members; they are also the basis for analogical errors in language learning and analogical changes in a language’s history. Changes such as the assimilation of wear : wered to the bear : bore pattern are legion, and they are easily explained if one assumes that the definition of a lexeme’s form paradigm is determined not by an inflection-class label but by principal parts and implicative relations. Principal parts do, however, raise an apparent problem of their own. Even in lightly inflected languages, paradigms very often admit alternative principal-part analyses; in English, for instance, the traditional analysis sing—sang—sung could instead have been singing—sang—sung or sings—sang—sung. In such instances, which analysis should be chosen as definitive? It is wrong, of course, to assume that there must always be a single, superior analysis among the range of available alternatives; it is also highly questionable to assume that speakers never store more forms than are minimally necessary to deduce the entirety of a lexeme’s paradigm. The principal-part approach does not depend on either assumption; it merely entails that speakers store enough of the forms in a lexeme’s paradigm for the remaining forms to be deducible, given the relevant network of implicative relations. Whether or not one chooses to exploit principal parts in the formal representation of a lexeme’s inflection-class membership, the concept of principal parts affords a new dimension for typological comparison. Suppose a lexeme L in some language has n cells c1 , . . . , cn in its paradigm. The power set of { c1 , . . . , cn } (i.e. the set of subsets of { c1 , . . . , cn }) is the set of candidate principal-part analyses for L; of these, some are viable, in that their realizations provide sufficient information to deduce the realization of L’s entire paradigm. The ratio of L’s viable principal-part analyses to its candidate analyses therefore reflects the predictability of L’s paradigm; the higher the ratio, the greater the predictability. As Finkel and Stump (2009) have shown, this set-theoretic measure of predictability reveals a good deal of variation both across languages and across inflection classes within a single language.14 14 See Moscoso del Prado Martín, Kostic, and Baayen (2004b) and Ackerman, Blevins, and Malouf (2009) for an alternative, information-theoretic measure of predictability that is not directly tied to the notion of principal parts; see also Blevins (this volume).

138

gregory stump

Typological comparison of principal-part systems reveals important limits on the complexity of inflection-class systems. The average number of principal parts required to deduce a given cell in a given paradigm is apparently always quite low. In their investigation of complex principal-part systems, Finkel and Stump (2009) find that in Fur (Nilo-Saharan; Sudan), the average number of principal parts required to deduce a given form in a verb’s paradigm ranges from 1.00 to 1.33 (depending on the verb’s conjugation), with the overall average across the nineteen conjugations being 1.04; in Comaltepec Chinantec (Oto-Manguean; Mexico), the average number of principal parts required to deduce a given form in a verb’s paradigm ranges from 1.00 to 1.75, the overall average across the sixty-seven conjugations being 1.08. These findings recall Cameron-Faulkner and Carstairs-McCarthy’s (2000) contention that there is a principled limit on the similarity of contrasting inflection classes, a limit which they formulate as the No-Blur Principle in (9).15 (9) The No-Blur Principle (Cameron-Faulkner and Carstairs-McCarthy 2000: 816) Among the rival affixes for any inflectional cell, at most one affix may fail to be a class-identifier, in which case that one affix is the class-default for that cell. This hypothesis has, however, proven to be too strong; the affixal expression of a lexeme’s inflection-class membership is not this consistently enforced. For discussion of evidence to this effect, see Stump (2005b; on Sanskrit) and Finkel and Stump (2009; on Fur).

6.9 Conclusion

.............................................................................................................................................................................

The notion of inflection classes has been recognized in grammatical theory for millennia; the rules in P¯an.ini’s grammar of Sanskrit, the As..ta¯ dhy¯ay¯ı (fourth century bce), make essential reference to the Dh¯atup¯a.tha, a list of nearly two thousand verb roots organized according to the roots’ conjugation-class membership. Even so, the precise status of inflection classes in a language’s grammar remains a matter of ongoing debate. This is the consequence of a number of interacting factors: •

Inflection classes are morphomic (Section 6.1) and are therefore inherently troublesome for grammatical theories that do not recognize the autonomy of morphology.

15

This principle is the descendant of Carstairs’ (1987) Paradigm Economy Principle, which CarstairsMcCarthy (1991: 222) formulates as in (i): (i) There can be no more inflectional paradigms for any word-class in any language than there are distinct “rival” inflectional realizations available for that morphosyntactic property-combination where the largest number of rivals compete.

inflection classes •

•

•

•

•

139

Actual inflection-class systems exhibit a range of deviations (both simple and complex) from the canonical ideal articulated by Corbett (2009b) (Sections 6.3, 6.4). The morphological expression of inflection-class membership is highly variable both within and across languages, being expressed in some instances by markings added to stems, in other instances by stem alternations, and in still other instances by both (Section 6.5). Inflection-class distinctions may (but need not) correlate with distinctions in phonology, syntax, and semantics, frequently in combination (Section 6.6); such correlations are often imperfect. Inflection classes evolve historically as an effect of a variety of causes (sound change, analogy, reanalysis); what is an inflection-class distinction at one stage in a language’s history may not be at another (Section 6.7). Inflection classes may be modelled theoretically in a number of ways, and they are the locus of considerable typological variation (Section 6.8).

chapter 7 ........................................................................................................

PARADIGMATIC DEVIATIONS ........................................................................................................

matthew baerman

7.1 Introduction

.............................................................................................................................................................................

Paradigmatic organization is one of the defining properties of inflectional morphology. For any given syntactic word class we assume that there is a common set of relevant features that its members need to inflect for. Morphology being what it is, the formal realization of these features may be far from straightforward (see Trommer, this volume), but our default assumption, as is clear from reading just about any syntactic description of a language, is that functional, morphosyntactic distinctions have corresponding formal, morphological distinctions. But real paradigms may deviate from this ideal picture. Sometimes multiple feature values map onto a single form (syncretism, Section 7.2). Sometimes they map onto the ‘wrong’ form (deponency, Section 7.3). And sometimes there is no form, or at any rate, not one that speakers will happily accept (defectiveness, Section 7.4). Roughly speaking, we have two interpretive choices here. On the one hand, we might try to minimize the apparent mismatch between form and function, and seek to reorganize our system of feature structure and semantics to more closely reflect the morphological paradigm. Alternatively, we may treat the deviations as morphological phenomena, whose motivation, if any, comes from the interaction of purely morphological elements. Either way, confronting these deviations forces a more nuanced view of morphological operations and the syntactic and semantic features that they realize.

7.2 Syncretism

.............................................................................................................................................................................

The paradigm in Table 7.1 shows the singular forms of a classifier verb from MurrinhPatha, the non-Pama-Nyungan Australian language described by Nordlinger (this

142

matthew baerman

Table 7.1 Singular forms of the verb ‘slash’ in Murrinh-Patha

1sg 2sg 3sg

nfut

p:ipfv

fut

fut:irr

pst:irr

pan than pan

nguni thuni puni

ngu thu pu

ngu thu ku

nguy thuy puy

Source: Nordlinger (this volume).

volume). Any account of this paradigm will need to include (at least) three persons and five distinct TAM values. Multiplied out, this should give us a paradigm of fifteen forms, but this one has only eleven. In some ways the patterns of syncretism make sense: the fact that future and future irrealis forms are conflated must surely have something to do with the fact that they both share the meaning future. But why this only affects the first and second person forms is less clear. And the conflation of 1sg and 3sg in the non-future is quite strange, in particular because the logic of the system appears to tell us that in the 1sg we have a 3sg form appearing in place of the expected *ngan. These conflations reveal a fundamental paradox underlying inflectional morphology and its relation to morphosyntactic functions. On the one hand it is inflectional forms that reveal to us the presence of morphosyntactic features and their values: the inventory of TAM, person, and number in Table 7.1 is itself derived from the alternations in form. In that respect we have to take the forms seriously. On the other hand, the presence of syncretism means that inflected forms may provide inconsistent testimony for the distinctions they are meant to encode. That is, inflectional morphology provides the linguist (and presumably the language user) with information that must sometimes be heeded and sometimes ignored. Similar mismatches are found in languages all over the world (Baerman and Brown 2005a, 2005b); while not a necessary property of inflectional paradigms, it is a frequent one. Sometimes a plausible explanation is indeed available directly in terms of the feature system. For example, in Cayubaba (a language isolate of Bolivia), shown in Table 7.2, verbs normally distinguish between first inclusive plural and first exclusive plural subject forms, but for the copula these are identical. An obvious analysis is to separate out the notions of person and clusivity, and say that ãreere marks only person and number, but not clusivity. The syncretic pattern might thus be taken as providing evidence that person in Cayubaba is a structured feature, and not simply a set of atomic values. This has been one of the major contributions of the study of syncretism to linguistic theory. Naturally, the success of such an approach depends on making a plausible case that there is indeed a natural class of shared semantic or functional properties. It may also require a notion of feature structure more elaborate than the simple hierarchy evident in Cayubaba, with the distinction of inclusive/exclusive subordinate to the first person plural. Consider the case paradigms in Table 7.3, from Matses, a Panoan language of Peru and

paradigmatic deviations

143

Table 7.2 Syncretism in the Cayubaba copula, showing plural forms

1incl 1pl 2pl 3pl

regular verb ‘jump’

copula ‘be’

tirasibi hirasibi pirasibie basibi

ãreere ãreere ãrepere ãreriki

Source: Key (1967: 30, 34).

Table 7.3 Case paradigms in Matses

abs erg gen

‘I’

‘we.incl’

‘who?’

ubi umbi cun

nuqui nuqui nuquin

tsuda tsuadan tsuadan

Source: Fleck (2003: 241).

Table 7.4 Feature decomposition of the syncretic forms in Table 7.3

abs erg gen

[–a,–b] [+a, –b] [+a,+b]

form

values

nuqui

[–b]

nuquin

[+a, +b]

Form

values

tsuda

[–a,–b]

tsuadan

[+a]

Brazil. The pronoun ‘I’ has distinct absolutive, ergative, and genitive forms, but in the other two pronouns the ergative is syncretic, either with the absolutive (in ‘we’) or with the gentive (in ‘who?’). If we attribute this pattern directly to shared feature values, then we need a structure in which the ergative can share something with both the absolutive and the genitive. Following an idea promoted by Jakobson (1936), we can decompose the cases into component values—provisionally [± a] and [± b]—and see the syncretic forms as realizing units of these components, as illustrated in Table 7.4. Alternatively, we can dispense with directly accounting the values conflated by the syncretic form, and treat it as an underspecified ‘elsewhere’ form. Consider the verb paradigms in Table 7.5 from Skou, a member of the same-named language family of

144

matthew baerman Table 7.5 Sample verb paradigms from Skou ‘call’

‘die’

singular plural

singular

1 2 ma 3.m 3.f

wa

‘jump’ plural

singular plural

wang

òe wing

ò

Source: Donohue (2004).

Table 7.6 Partial paradigm of the verb ‘eat’ in Chipaya

1excl 1incl 2 3

Present absolutive

Future

lul-u-tra

lula-tra

lul-tra

lula-ki-tra

Source: Cerrón-Palomino (2006: 148f).

New Guinea. Each paradigm involves two forms: one is fully specified for all the relevant subject features (person, number, gender), and the other takes care of the rest of the paradigm. The opposition between overtly specified and elsewhere forms is often equated with the distinction between overtly marked and unmarked forms, as in the obvious case of English s/he walk-s vs. I/we/you/they walk. ‘Obvious’ can be deceptive though. Consider the paradigms in Table 7.6, from Chipaya, one of the two constituents of the Uru-Chipaya language family of Bolivia. In the present absolutive there is an overt -u suffix marking first person (exclusive), opposed to an unsuffixed form which covers the remaining values, namely second and third person, as well as first person inclusive (the -tra which follows is a declarative suffix).1 This looks like a fairly clear case of syncretism due to morphological unmarkedness. However the future displays exactly the same paradigmatic pattern, but with the morphological markedness reversed: the non-first person form is suffixed, and the first person form is unsuffixed. Most approaches to syncretism involve some kind of juggling between morphosyntactic natural classes and underspecification. There are, however, some paradigm 1 Which itself is sensitive to the gender and age status of the discourse participants; -tra presupposes an adult speaker and a male listener (Cerrón-Palomino 2006: 166).

paradigmatic deviations

145

Table 7.7 Paradigm of the verb ‘work’ in Skou

1 2 3m 3f

Singular

Plural

lóe póe lóe póe

tóe lóe tóe tóe

Source: Donohue (2004).

Table 7.8 Varieties of alternations in Dhaasanac (perfective forms) ‘walk’

‘die’

singular plural singular

‘kill’ plural

singular plural

1incl ——

seð

——

kuufumi

——

yes

1

seð

sieti

kuufumi

kuufunanni

yes

ces

2

sieti

sieti

kuufunanni

kuufunanni

ces

ces

3.f

sieti

seð

kuufunanni

kuufumi

ces

yes

3.m

seð

seð

kuufumi

kuufumi

yes

yes

Source: Tosco (2001: 146, 173, 200).

configurations which defy such treatment. Consider another example from Skou, the paradigm of the verb ‘work’. There are three forms: lóe in the 1sg, 2sg masculine, and in the second plural, póe in the 2sg and 3sg feminine, and tóe in the 1pl and 3pl. None of these can readily be reduced to a morphosyntactic natural class, nor can they all be treated as defaults: an elsewhere form can only be understood as standing in opposition to an explicitly defined form. If we want to represent such patterns in a morphological description, the set of values that have been formally conflated must be overtly stipulated. To do so means giving morphology greater autonomy than many linguists are willing to accord it, so the condition ‘if we want to represent such patterns in a morphological description’ plays a crucial role. Where a pattern is robustly attested across multiple contexts, doing so is unavoidable, or at least nearly so. For example, in Dhaasanac (a Cushitic language of Ethiopia) the verbal paradigm contains two subject forms: one for the second person, first person exclusive plural, and third person feminine singular, and one for first person singular, first person plural inclusive, third person masculine singular, and third person plural (see Table 7.8). The morphosyntactic values cannot be conflated, but the pattern is quite regular and consistent, and is morphologically realized in various different ways (suffixation, stem-final consonant alternations,

146

matthew baerman Table 7.9 Patterns of person syncretism in verbal subject marking in Kaluli, Uduk, and Ayutla Mixe

1 2 3

Kaluli ‘see’

Uduk (Koman) ‘eat’ (transitive ipfv)

Ayutla Mixe ‘work’

bodo:l bodo:l badab

shwana shwa shwa

tunp mtunp tunp

Sources: Kaluli: Grosh and Grosh (2004: 102); Uduk: Killian and Hammarström (2009: 10); Ayutla Mixe: Romero-Méndez (2008: 296).

ablaut, prefixation). Since multiple morphological operations make reference to the same pattern, if that pattern cannot be derived, it must be morphologically stipulated. However, such clear cases of morphological systematicity are in the minority. Indeed, most examples of syncretism are ambiguous, so it is seldom clear which of the three analyses outlined above (morphosyntactic identity, underspecification, morphological stipulation) is the most appropriate. There is undoubtedly a great deal of variation in the patterns of syncretism attested in different languages, but cross-linguistically we can observe certain conflations that occur over and over, particularly involving person marking and core case values (which may itself simply be due to the fact that these features themselves are relatively consistent cross-linguistically). Table 7.9 shows examples of various patterns of person syncretism, involving subject agreement on verbs, in paradigms where there is no number distinction. By far the most common patterns are for first and second person to be conflated, as in Kaluli (a Trans-New Guinea language of the Bosavi family), or second and third, as in Uduk (a Nilo-Saharan language of the Koman family, spoken in South Sudan). Conflation of first and third, as in Ayutla Mixe (a Mixe-Zoque language of Mexico), is less common (Baerman et al. 2005: 59–61). This suggests a semantic affinity of first and second person (often classed as speech act participants) and second and third (under the less attractive label non-first person, although note that some pronominal systems do instantiate such a unit; see Cysouw 2003), alongside the absence of any such affinity between first and third person. In case syncretism, cross-linguistically common patterns involve the interaction between core cases, and their relationship to non-core cases, if we understand core cases to be those used for the basic syntactic functions of intransitive subject, transitive subject, and direct object—that is, various configurations of nominative, absolutive, ergative, and accusative. The basic patterns are illustrated in Table 7.10, using the Matses paradigms repeated from Table 7.3, and some partial paradigms from Polish. Matses represents an ergative-absolutive system and Polish a nominative-accusative system; the non-core cases are represented by the genitive (the only non-core case

paradigmatic deviations

147

Table 7.10 Alternating patterns of syncretism in an ergative and accusative system Matees

abs erg gen

Polish

‘I’

‘we.incl’

‘who?’

ubi umbi cun

nuqui nuqui nuquin

tsuda tsuadan tsuadan

nom sg acc sg gen sg

‘book’

‘knife’

‘student’

ksi˛ez˙ k-a ksi˛ez˙ k-˛e ksi˛ez˙ k-i

nóz˙ nóz˙ noz˙ -a

student student-a student-a

in Matses, one of a few in Polish). Both languages display core case syncretism: absolutive/ergative in Matses (‘we’), nominative/accusative in Polish (‘knife’). This by far the most common type. Less common, but still widely attested, is the pattern in which one of the core cases is syncretic with a non-core case: the ergative in Matses and the accusative in Polish. In particular, it looks as if the oblique case form is used for the marked core case (where marked is understood as morphologically marked, with absolutive and nominative typically being morphologically unmarked). That is, in Matses the suffix -n generally characterizes the genitive, but is also found in the ergative in ‘who’, and in Polish, -a characterizes the genitive singular in one inflection class, but is found in the accusative singular too of such items as ‘student’. Both the syncretism of core cases and the syncretism of the marked core case with a non-core case have a likely explanation, at least at some point in their development, in the notion of differential argument marking (Bossong 1991). In such a system, the marked core argument is only distinguished when it fulfils certain semantic, syntactic, or discourse related conditions, such as animacy, word order, or definiteness. For example, in Matses, singular (or at any rate, not overtly plural) pronouns distinguish ergative and absolutive, while plural pronouns do not. In Polish, within certain inflection classes animates distinguish accusative and nominative but inanimates do not (‘student’ vs. ‘knife’; ‘book’ belongs to a different inflection class). The pattern wherein the non-core case form is used for the marked core case appears to arise under these conditions when no dedicated core case form is available. Typical mappings are the use of an instrumental or genitive as an ergative, or a genitive or dative as an accusative.

7.2.1 Syntactic Effects of Syncretism In some instances syncretism creates a set of feature values that can be exploited by syntactic rules in the resolution of conflicts in agreement and government (Pullum and Zwicky 1986; Corbett and Mtenje 1987). For example, in Icelandic, dative subjects take default third person agreement, and any accompanying nominative object must normally be third person as well (1a, b), so that person agreement is in effect

148

matthew baerman

doing double duty here. In the singular preterite paradigm, first and third person are syncretic (as they are throughout Germanic), and surprisingly, first person singular nominative objects are found to be at least marginally acceptable (1c). That is, syncretism enables what would otherwise be a syntactically illicit construction. Conversely, syncretism may have a deliterious effect, leading to gaps in the paradigm (see Section 7.4.1.2 below). (1)

a.

Henni líkað-i her.dat liked-1/3.sg ‘She liked him.’

hann he.nom

b.

*

Henni líkað-ir her.dat liked-2.sg ‘She liked you.’

þú you.nom

c.

??

Henni líkað-i her.dat liked-1/3.sg ‘She liked me.’

ég I.nom (Sigurðsson 1996: 33, cited by Wood 2010)

7.3 Deponency

.............................................................................................................................................................................

Like syncretism, deponency involves a many-to-one mapping between meaning and form, but across paradigms rather than within a paradigm. The eponymous examples come from Latin, illustrated in (2). Transitive verbs display a regular contrast between active and passive forms, as in active amant ‘they love’ vs. passive amantur ‘they are loved’ in (2). A deponent verb has the same forms as the regular passive, but is syntactically and semantically active; thus the deponent hortantur ‘they exhort’ in (3) is an ordinary active transitive, not a passive. (2) qu-ae ex se nat-os ita am-ant ad quoddam which-n.pl from self.abl born-acc.pl thus love-3pl to certain.acc.sg tempus time.acc.sg

et and

ab from

e-is them-abl.pl

ita thus

am-antur love-3pl.passive

‘. . . which [animals] thus love their offspring for a certain time and thus are loved by them.’ (Cicero, De amicitia, Chapter VIII) (3) me=que hort-antur ut magn-o anim-o sim me.acc=and exhort-3pl that great-abl.sg spirit-abl.sg be.1sg.sbjv ‘. . . and they exhort me to be of good courage’ (Cicero, Epistulae ad Atticum, book 11, letter 6)

paradigmatic deviations

149

What is striking about the Latin deponents is that they have the formal trappings of a systematically marked grammatical category without having the associated meaning. But if passive morphology ‘means’ passive—which it does productively for the vast bulk of transitive verbs—then how do we account for the fact that it sometimes has quite an incompatible, indeed opposite, meaning? If we conceive of inflectional forms as realizing morphosyntactic values, how do we allow the production of forms in the absence of those values? Questions of this sort arise not infrequently when looking at inflectional systems, and borrowing the term deponency from Latin grammar is a convenient way of talking about it. However, to make it a useful typological notion requires separating certain general properties from the specific peculiarities and difficulties of the Latin example. A general definition is offered in (4). (4) Informal definition of deponency In one paradigm there is a set of forms ‘X’ that have the morphosyntactic value x. In another paradigm we also find a set of forms ‘X’, but here they have the value y (where y ̸= x). Note that this definition makes reference not to a form, but to a set of forms. This is not because it is conceptually central; rather, it ensures that what we are observing is not an instance of accidental homophony. For example, the Latin passive differs from the active in a variety of different ways, some examples of which are shown in Table 7.11, through various suffixes (-r, -ur, -re, -mini, -¯ı), infixes (-ri-) or periphrasis, as in the perfect forms. These are faithfully matched in the paradigm of deponent verbs. This demonstrates that whatever it is that connects the two paradigms, it is morphologically systematic. Although the definition in (4) makes no reference to the actual morphosyntactic categories involved, there are many examples which do involve argument structure.

Table 7.11 Representative sample of the active ~passive/deponent opposition normal verb ‘love’

prs ind 1sg prs ind 2sg prs ind 3sg prs ind 1pl prs ind 2pl prs ind 3pl imp sg inf prf ind 1sg

active

passive

deponent verb ‘exhort’

am¯o am¯as amat am¯amus am¯atis amant am¯a am¯are am¯av¯ı

am¯or am¯aris am¯atur am¯amur am¯amin¯ı amantur am¯are am¯ar¯ı am¯atus sum

hortor hort¯aris hort¯atur hort¯amur hort¯amin¯ı hortantur hort¯are hort¯ar¯ı hort¯atus sum

150

matthew baerman

Table 7.12 Number deponency in Serbo-Croatian ‘child’

nom acc gen dat ins

neuter et-stem ‘button’

a-stem ‘woman’

singular

plural

singular

plural

singular

plural

dete dete detet-a detet-u detet-om

dec-a dec-u dec-e dec-i dec-om

dugme dugme dugmet-a dugmet-u dugmet-om

dugmet-a dugmet-a dugmet-a dugmet-ima dugmet-ima

žen-a žen-u žen-e žen-i žen-om

žen-e žen-e žen-a žen-ama žen-ama

For example, Bantu languages are often described as having pseudo-passives, pseudocausatives, pseudo-applicatives, etc., which have the morphological characteristics of these categories without the associated meaning or argument structure (see Good 2007, who does express doubts about many putative instances). Examples involving other grammatical categories are harder to come by, but are by no means absent. An example with nouns from Bosnian-Croatian-Serbian is shown in Table 7.12, where the plural of the word for ‘child’ has what appear to be singular forms. Although its paradigm contains only singular forms, there is still a morphological contrast between singular and plural, on account of the heteroclitic pattern (Stump 2006, this volume), namely the juxtaposition of forms drawn from two distinct paradigms. Thus its singular forms are those of a neuter noun belonging to the stem class that takes an et-extension in most of the paradigm, such as ‘button’. Its plural forms correspond to the singular forms of an entirely different inflection class, the a-stems, as represented here by ‘woman’. Agreement patterns testify to the inherent tension within the plural paradigm of dete, with both singular and plural agreement being found, according to the Agreement Hierarchy (Corbett 1983: 85). The example in Table 7.12 reveals a further element of typological variation: if a deponent item uses form x to realize value y, how does it realize the now orphaned value x? The deponent verbs of Latin, having used up their options for marking the passive, simply lack passive forms and are defective (if transitive). The BosnianCroatian-Serbian example uses heteroclisis to flesh out the paradigm. A third possibility is that the feature values involved trade places, so that the paradigm of a deponent lexeme would be the mirror image of a normal lexeme, corresponding to the phenomenon often known as polarity (Hetzron 1967). For example, in Gulmancema (a Gur language of Burkina Faso), the distinction between perfective and imperfective may be marked by suffixation in one of the aspects (di, -li, -gi, or -ni) or an alternation between high and low tone. In the first set of verbs in Table 7.13 it is the imperfective which is marked by suffixes (a–d) or by high tone. In the second set of verbs we see exactly the same morphological operations, but with the reverse function.2 In effect, 2

Note that not every type of morphological opposition between perfective and imperfective is reversable in this way; see Naba (1994: 356–80).

paradigmatic deviations

151

Table 7.13 Aspectual alternations in Gulmancema

a. b. c. d. e.

pfv

ipfv

ba ba bia sagi pwà

baa-di baa-li bia-gi sag-ni pwá

‘get’ ‘fall’ ‘get worse’ ‘fry’ ‘hit’

pfv

ipfv

jii-di bu-li ŋma-gi gua-ni dá

ji bu ŋmia gua dà

‘go down’ ‘sow’ ‘twist’ ‘sleep’ ‘buy’

Source: Delplanque (2009).

each row in Table 7.13 represents a single morphological opposition which maps onto two types of paradigm.

7.3.1 Interpretations As with syncretism, deponency can be taken as revealing something about the structure of morphosyntactic features, or about the morphological operations that realize these features. On a feature-based approach (e.g. Embick 2000; Kiparsky 2005), deponency is treated as an instance of the lexical assignment of morphosyntactic features. In its role as a lexically assigned property, a feature value such as passive brings about all the surface effects of the passive—the appropriate inflected forms, as well as their participation in periphrastic constructions—without the associated interpretation. A morphology-based approach (Sadler and Spencer 2001; Stump 2006) treats deponency as evidence for the split between a morphosyntactic paradigm and a form paradigm. While the default situation is for there to be a one-to-one mapping between them, by separating the two it is possible to model a many-to-one mapping between morphosyntactic and form paradigms, for example with the passive of ordinary transitive verbs and the active of deponent verbs both mapping onto the same form paradigm. The two types of approach imply different interpretations, which are of broad relevance beyond the particulars of individual formal analyses. The feature-based approach implies a fundamental asymmetry: the natural environment of a morphosyntactic is grammatical, its use as a lexical feature is exceptional. A morphologybased approach implies a certain symmetry: although one of the alternative mappings can be flagged as the default, the ‘normal’ and the deponent paradigms involve morphological rules of the same sort. These implicit distinctions have a parallel in descriptive conventions. The standard account of Latin treats productive passive formation first, and then describes deponent verbs by reference to this. Contrast this with Munro and Gordon’s (1982) account of Chickasaw, a Muskogean language originally

152

matthew baerman

spoken in parts of the southeastern United States. Intransitive verbs display what is sometimes called split intransitivity: some intransitive verbs mark their subjects with the same affixes as are used for the subject of a transitive verb (compare (5a) and (6a)), others use the same affixes that are used for the object of a transitive (compare (5b) and (6b)).3 (5)

(6)

a.

a.

ila-li different-1sg ‘I am different.’

b.

chi-hoyo-li 2-look.for-1sg ‘I look for you.’

b.

sa-chokma 1sg-good ‘I am good.’

(Munro and Gordon 1982: 84f)

is-sa-hoyo 2-1sg-look.for ‘You look for me.’

(Munro 2005: 125)

Although on the whole the lexical split in the intransitives largely corresponds to a semantic split between agentive and patientive subjects, Munro and Gordon (1982: 84) argue that there is no predictable mapping from semantics to morphology, and that for most verbs the affixation pattern is fixed even where the semantics of the expression varies. Thus, they construe the distinction between the two types of intransitive in (5a) and (5b) as lexically specified. Given this analysis, one could treat the form such as (5b) as a mismatch: the subject is marked like an object. This is not Munro and Gordon’s approach; instead, they set up a purely morphological distinction between set I and set II affixes, and then describe how the grammatical functions map onto it: transitive subjects and some intransitives take set I affixes, objects and other transitives take set II. Thus on a feature-based approach, deponency is treated as a genuine mismatch: there is a lexically-specified feature whose value does not correspond to its actual morphosyntactic function. On a morphology-based approach, there is no mismatch, because there are no privileged matches: rather, such phenomena are taken as evidence of a fundamental split between morphological form and morphosyntactic function.

7.4 Defectiveness

.............................................................................................................................................................................

A defective paradigm is one lacking one or more of the forms which it should be expected to have by any reasonable non-morphological criteria. A familiar example from Russian is shown in Table 7.14. The normative tradition states that the word meˇcta ‘dream’ lacks a genitive plural, even though the behaviour of other members of its inflection class make it fairly clear what the form ought to be (namely, the bare stem). Indeed, corpus data from the Russian National Corpus (RNC) shown in Table 7.14 confirm that the genitive plural is vanishingly rare: out of 8,722 tokens of this lexeme, 3 And further, there are intransitive verbs whose subject is marked in the same way as indirect objects; indirect object markers are the same as the direct object markers with the addition of a final nasal.

paradigmatic deviations

153

Table 7.14 Russian ‘dream’

nom acc gen dat loc ins

a. normative paradigm

b. tokens in the RNC (queried 10/2008)

singular

singular

meˇctá meˇctú meˇctý meˇcté meˇctój

plural meˇctý *meˇct meˇctám meˇctáx meˇctámi

2771 929 996 362 643

plural 1980 13 221 459 348

Note: Forms that are syncretic within the singular or plural have been conflated.

only thirteen are of the genitive plural, and most of these are clearly jocular or ironic, playing on the apparent incongruity of the form. Crucially, there is nothing about the meaning of the word that would account for this gap—unlike, say, the absence of plural forms for a non-count noun. Nor is there anything phonotactically objectionable about the expected form. Nevertheless it appears to be widely avoided. In its place, various alternative strategies are employed, such as using a different lexeme4 or a non-canonical syntactic construction,5 or framing the sentence so as to avoid any context which would call for the genitive plural. Seen from the perspective of the architecture of grammar, defectiveness represents an unwanted intrusion of morphological idiosyncrasy into syntax. From a morphology-internal perspective it is also problematic, since most models of inflectional morphology assume that there will always be productive, default mechanisms available to generate inflected forms, which must be suspended in the case of defective items. Although defective paradigms are periodically noted in language descriptions, and indeed have been incorporated into the normative grammars of a number of European languages, for obvious reasons they are difficult to investigate: it may be only serendipity that reveals them in the first place, and once identified, judgements of acceptability are gradient and vary from speaker to speaker. Nor is there any reason to think that defectiveness represents in any way a unified phenomenon. Any number of factors may contribute to the failure of a paradigm. Nevertheless, most observed examples can fit within a broad typological categorization which points in the direction of likely explanations. 4

Such as meˇctanie ‘dreaming, dream’, whose genitive plural is unusually frequent: 628 tokens (51%) out of a total 1,230 for the lexeme as a whole in the RNC (queried 10/2008). Presumably it is taking some of the load off meˇcta. 5 For example, there are some quantifiers which normally take the genitive plural of count nouns (such as mnogo ‘many’, or the oblique forms of the numerals ‘2’, ‘3’, and ‘4’), but the RNC contains examples of these quantifiers used with the genitive singular of meˇcta.

154

matthew baerman Table 7.15 Terms for musical notes in Turkish ‘fa (note F)’ default form 1sg -Vm 1pl -VmVz

fa *fa-m fa-mız

‘sol (note G)’ ‘F’ ‘my F’ ‘our F’

sol sol-üm sol-ümüz

‘G’ ‘my G’ ‘our G’

Source: Inkelas and Orgun (1995: 770).

7.4.1 Typology The typology offered here is based on the linguistic components that need to be invoked in order to describe the paradigmatic gaps, broadly divided into (i) morphophonology, (ii) morphological form, and (iii) morphosyntactic features. Such a purely descriptive typology may in turn have ramifications on the interpretation; for example a gap that can be described in morphophonological terms is probably not the same sort of thing as one that can only be described in terms of morphosyntactic feature values, since the former may ultimately be related to phonological constraints, and the latter to syntactic or semantic constraints.

7.4.1.1 Morphophonology Inkelas and Orgun (1995), citing their own work as well as Itô and Hankamer (1989), report that some speakers of Turkish observe a disyllabic minimal size condition on affixed words. In practice, this means a ban on forms involving a monoconsonantal suffix on a monosyllabic vowel final stem. Although where this situation obtains for the native lexicon, alternative morphological strategies are employed, real problems do arise with the names of letters of the alphabet and musical notes with the shape CV when combined with the first singular (-Vm) and second person singular (-Vn) possessive suffixes, as the initial vowel of the suffix is regularly deleted when attached to a vowel-final base. Note that (i) disyllabic possessive suffixes are acceptable, and (ii) monosyllabic suffixes are acceptable so long as the stem is consonant-final, as the resulting form ends up disyllabic.

7.4.1.2 Morphological form Probably the most striking examples of defectiveness are those where the gap corresponds to some element of morphological form, because this is precisely where a semantic or phonological account would be hard to maintain. The morphological elements can be broadly characterized as involving: (i) morphological alternants, (ii) morphotactics, and (iii) paradigms. First, a paradigmatic gap may correspond to the absence of an expected morphological alternant, in particular a stem alternant. Consider the abbreviated paradigm of the

paradigmatic deviations

155

Table 7.16 Defective verb ‘spread out’ in Finnish finite

ind prs ind pst cond prs pot prs 2sg imp 3sg imp

active

impersonal

erkanee erkani erkanisi ——— erkane ———

——— ——— ——— ——— ———

non-finite prs ptcp pst ptcp 1st inf 2nd inf 3rd inf

erkaneva ——— ——— ——— erkanemaan

Note: Finite active forms are 3sg unless otherwise noted

Table 7.17 Non-defective verb ‘be quiet’ in Finnish finite

ind prs ind pst cond prs pot prs 2sg imp 3sg imp

active

impersonal

vaikenee vaikeni vaikenisi vaienee vaikene vaietkoon

vaietaan vaietiin vaiettaisiin vaiettaneen vaiettakoon

non-finite prs ptcp pst ptcp 1st inf 2nd inf 3rd inf

vaikeneva vaiennut vaieta vaietessa vaikenemaan

defective Finnish verb ‘spread out’ in Table 7.16. It is missing an odd set of forms: the potential present, the 3sg imperative, the past participle, first and second infinitive, and all the impersonals (see Kiparsky 1974; Haarala 1996; Hakulinen et al. 2004).6 But if we look at the paradigm of a normal verb of the same inflection class, such as ‘be quiet’ (Table 7.17), we see an obvious morphological correspondent to the gap in ‘spread out’. The non-defective verb has two stems, the strong grade vaikene- and the weak grade vaie-; the gap in the paradigm of ‘spread out’ corresponds to the range of the weak grade stem.7 Thus to describe the defectiveness in ‘spread out’, we can say that the verb lacks a weak grade stem, and consequently all the forms that are based on it. This sort of stem-based gap accounts for some of the better known examples of defectiveness, being well represented in Romance languages (in particular, French and Spanish); see Maiden (1992) and Boyé and Cabredo-Hofherr (2006). 6 Roughly synonymous verbs based on the same root but belonging to different verb classes, such as erota or erkaantua, are said to be used instead. 7 Note that verbs of this type normally have a base in -e-, e.g. vaikenee ‘is quiet’, whereas the base of erkanee exceptionally has -a.

156

matthew baerman Table 7.18 Agreement markers in Chickasaw, typical mapping for transitive verbs

1sg 1pl 2 3

set I (subject)

set II (object)

-li (k)iiishØ

sapochiØ

Source: Munro (2005: 124f).

Second, the gap may correspond to an otherwise expected combination of morphological formatives. Chickasaw, discussed in Section 7.3.1, offers an example. Recall that verbal argument markers are divided into set I and set II, and that transitive verbs take set I markers for their subject and set II markers for their object. That is the normal case, but there is a small set of transitive verbs, such as banna ‘want’ and nokfonkha ‘remember’, that take set II markers for their subjects instead of set I (Munro 2005: 125). But if these verbs take object markers for their subjects, how are objects treated— are they drawn from set I or set II? In the case of third person arguments the answer is easy, as these are not overtly marked in any case (see Table 7.18). Thus a form like sa-banna ‘I want her’ has a set II marker for the subject, and its third person object is implied by the lack of further affixation. But according to Munro (2005: 125), speakers are apparently unwilling to produce forms with first or second person objects, resorting instead to some paraphrase. Interestingly, speakers of closely related Choctaw apparently have no objection to this, and the corresponding verbs simply have two set II markers, one for the subject and one for the object (7). Thus the gap in Chickasaw can be described in terms of morphotactics, namely a constraint against the co-occurrence of two set II markers in single word form. (7)

An-at-o 1sg-nom-cont ‘I want you.’

chi-sa-banna-h 2-1sg-want-predicative (Davies 1986: 67; see also Broadwell 2006: 145)

Third, some instances of defectiveness seem to involve reference to the rest of an inflectional paradigm. For example, the genitive plural gap in Russian that introduced this section (Section 7.4) can be described, at least roughly, by reference to the accentual paradigm of the noun (Es’kova 1989: 98; see also Pertsova 2005). Nouns of this inflection class display one of three stress patterns, exemplified in Table 7.19: fixed stress on the stem (‘mast’), stress which alternates between stem and ending (‘water’), and fixed stress on the ending (‘dream’). But because the genitive plural in this class has a null ending, stress here necessarily falls on the stem, which means that the accentuation of the genitive plural in the otherwise ending-stressed third type is anomalous. Very few

paradigmatic deviations

157

Table 7.19 Accentuation of Russian class II nouns

nom acc gen dat loc ins

fixed stem stress

alternating stress

fixed ending stress

‘mast’

‘water’

‘dream’

singular

plural

singular

plural

singular

máˇct-a máˇct-u máˇct-y máˇct-e máˇct-e máˇct-oj

máˇct-y máˇct-y máˇct máˇct-am máˇct-ax máˇct-ami

vod-á vod-ú vod-ý vod-é vod-é vod-ój

vód-y vód-y vód vod-ám vod-áx vod-ámi

meˇct-á meˇct-ú meˇct-ý meˇct-é meˇct-é meˇct-ój

plural meˇct-ý meˇct-ý *méˇct meˇct-ám meˇct-áx meˇct-ámi

Table 7.20 Syncretism versus defectiveness in perfective statives in Tuareg

1sg 2sg 3sg 1pl 2pl.m 2pl.f 3pl.m 3pl.f

Tamahak

Tamashek

karrozʕ əɣ karrozʕ əd karrozʕ karrozʕ karrozʕ əm karrozʕ mət karrozʕ ən karrozʕ nət

kæwɑlæɣ kæwɑlæd kæwɑl ——– kæwɑlæm kæwɑlmæt kæwɑlæn kæwɑlænt

Sources: Tamahak: Cortade (1969: 41); Tamashek: Heath (2005: 437f).

nouns are of this type, but defective meˇcta is one of them, and by far the most frequent. This gap can be described in terms of the relationship between forms in the paradigm. An interesting variant of this occurs in apparent instances of homophony avoidance, which involves the intersection of syncretism and defectiveness (see also Hansson, this volume). For example, in the Tamashek of Mali (one variety of Tuareg, a Berber language—or perhaps group of languages) described by Heath (2005), the inflectional rules for the perfective of stative verbs would seem to generate a first person plural form homophonous with the third person singular, both of them represented by the bare stem (perfective formation in these verbs appears to involve the deletion of the subject prefixes found in the rest of the paradigm) Such a form has indeed been reported for Tamahak, another variety of Tuareg, spoken in Algeria (see Table 7.20; Cortade 1969: 41). But Heath reports that his Tamashek consultants rejected this form for the first

158

matthew baerman Table 7.21 Witsuwet’en normal verb ‘breathe’ s#giz ∗

1 2 3

defective ‘walk fast’ n#wət

singular

dual

plural

singular

dual

plural

s-əs-giz s-en-giz sgiz

s-dət-giz s-əxw -giz hə-sgiz

ts’ə-sgiz

——– ——– nəwət

——– ——– hə-nəwət

——–

Note: ∗ Hargus gives the underlying form as the abstract D-yiz, from which s-giz can be derived by regular morphophonological rules. Source: Hargus (2007: 777, 569).

person plural, and in fact showed no general agreement about how to produce a form with this meaning, some even resorting to paraphrase.

7.4.1.3 Morphosyntax Finally, there are some gaps that seem not to correlate with any units either of form or meaning, and can only be described by listing the morphosyntactic feature values involved. For example, in Witsuwet’en (an Athabaskan language of British Columbia), Hargus (2007: 569) reports that the verbs n-wət ‘walk fast’ and c’-ɬ-tsi ‘give birth’ are used only in the third person singular and plural (see Table 7.21). There is no obvious semantic justification—while a restriction to 3rd person subjects might be unsurprising in the case of verbs with inanimate or impersonal subjects, these verbs necessarily have animate subjects. And the gaps may be filled by alternative, semantically equivalent expressions, e.g. ʔaγ nə-s-ye ‘fast walk’ in place of *nə-s-wət ‘walk.fast’ (Hargus 2007: 569). Likewise, there is no morphological correlate: for example the missing forms do not correspond to an expected stem alternant, and while the affixes can be divided into two sets according to position class, this split does not correspond to the observed gap . The only generalization that captures the defective pattern is to say that these verbs only have forms for thrid person subjects.

7.4.2 Explanations Defectiveness is by definition exceptional and dysfunctional, so it is natural to think that it must be due to some kind of inherent oddity in the paradigm. Certainly, for the types described in the typology above under morphophonology (Section 7.4.1.1) and morphology (Section 7.4.1.2), it is generally the case that the forms are in some way

paradigmatic deviations

159

aberrant; or if not the forms themselves, their relationship to the rest of the paradigm. Albright (2003) has suggested that gaps in such instances are driven by a lack of confidence; that is, even where the inflectional rules might seem to leave little doubt about what the form ought to be, there are too few examples of the rule being applied in the relevant context to be absolutely sure. Whatever the underlying or original motivation, there may also be an element of arbitrary conventionalization. Daland et al. (2007) have even offered evidence that language learners can deduce and propogate gaps based simply on (in-)frequency.

7.5 Conclusion

.............................................................................................................................................................................

Apparent aberrations in paradigmatic organization provide a useful yardstick for evaluating approaches to inflectional morphology. Given a particular set of morphosyntactic features, any deviation from a simple mapping between these and actual morphological forms will require some elaboration of the model, targeting either the features or the rules. There is and always has been a lively debate about what sorts of elaborations should be considered—in particular, how much work, if any, should be given over to morphological stipulation. Deviation in the sense discussed here is thus not an absolute identification of some inherent perversity, but rather an indication that the model needs fine tuning.

Acknowledgements

.............................................................................................................................................................................

The writing of this chapter was funded by the European Research Council under grant ERC-2008-AdG-230268 MORPHOLOGY. Their support is gratefully acknowledged.

chapter 8 ........................................................................................................

PHONOLOGY ........................................................................................................

gunnar ólafur hansson

8.1 Introduction

.............................................................................................................................................................................

The interface between the subsystems of phonology and morphology (inflectional and derivational alike) remains a productive area of research and debate within theoretical phonology. The advent of constraint-based frameworks within generative phonology, in particular Optimality Theory (Prince and Smolensky 2004), and the general shift of the theoretical mainstream from serialist transformation-based models to (largely) parallellist constraint-based models, has had a considerable impact on current thinking with respect to morphology–phonology interactions. For example, the strong drive toward parallel, global, and monostratal computation in (standard) Optimality Theory puts into sharp focus any phenomena that have traditionally suggested the need for serial, cyclic, or multistratal derivations: ‘cyclicity effects’, construction-specific phonology, and the like. Furthermore, the formal notion of correspondence, and the possibility of extending correspondence constraints beyond the input–output dimension (underlying to surface representation) to the output–output dimension (relations among surface forms) has invited the explicit formalization of such traditional notions as analogical levelling and paradigm uniformity within a generative framework. Finally, the output-oriented character of Optimality Theory, in which all phonological computation is seen as inherently optimizing, has prompted a re-evaluation of many areas in which morphological processes are subject to phonological conditioning (e.g. allomorph selection, as well as other types of ‘unusual’ exponence; see Trommer and Zimmerman, this volume). This re-evaluation has gone hand-in-hand with a renewed focus on cross-linguistic typology, from which the field has benefited greatly. This chapter will first briefly review current approaches to modelling the morphology–phonology interface, and the mechanisms by which information about morphological structure can affect phonological patterns (Section 8.2). We then turn

162

gunnar ólafur hansson

to two topics of particular interest with respect to the mutual interaction of morphology and phonology. The first of these is the interplay of paradigms and paradigm structure with phonology (Section 8.3). This includes the levelling of phonological alternations (paradigm uniformity effects), the blocking or triggering of phonological processes due to homophony avoidance (paradigm anti-homophony effects), and the occurrence of paradigm gaps which appear to be phonologically motivated. Secondly, we will examine in detail the phenomenon of phonologically conditioned suppletive allomorphy (Section 8.4), in which the direction of influence is from phonology to morphology rather than the reverse. The morphology–phonology interface is an expansive and complex domain of inquiry, and many of the central theoretical questions cut across the divide between inflection and word formation. This chapter makes no pretence at exhaustive coverage, but rather focuses on a few theoretical issues and empirical phenomena which are of particular relevance for inflectional morphology. For other recent handbookchapter overviews of the morphology–phonology interface in general, or of specific topics addressed within the present chapter, see Albright (2011), Bermúdez-Otero (2011, 2012), Inkelas (2011), Kaisse and McMahon (2011), Nevins (2011), Trommer (2011b), and Trommer and Zimmerman (this volume).

8.2 Modelling the Morphology–Phonology Interface ............................................................................................................................................................................. The phonology of a language L can be thought of as a function ϕL from an input to an output representation: ϕL (X) = Y thus denotes that in language L an input (lexical, underlying, phonemic) representation /X/ surfaces as [Y]. (Henceforth the subscript L will be omitted in the interest of typographical clarity.) In current generative models ϕ not only encodes allophonic detail but can also cover various neutralizing (featurechanging) processes, deletions, insertions, and prosodic structure-building such as syllabification, stress assignment, and so on.1 Similarly, we can take morphology to be a function µ, which can at least in some cases be seen as composed of a series of nested functions, µ( · ) = µn (µn−1 ( . . . (µ2 (µ1 ( · ))))), where each µi represents one derivational step or operation (e.g. affixation). There are various ways in which one might conceptualize the manner by which phonology and morphology interact, such that phonological patterns and processes can be sensitive to aspects of morphological structure or vice versa. At one extreme, we might imagine that all morphology logically precedes all phonology in the sense that the input to the phonological grammar is a fully structured (and fully ‘spelled-out’) 1 Note that this formulation abstracts away from the question of how ϕ is best modelled formally— L for example, in terms of serially ordered transformational rules, or ranked and violable output constraints—which is of course a matter of central importance to phonological theory.

phonology

163

object which contains, in addition to an underlying phonological representation, all details regarding the internal morphological make-up of the word form in question (constituent structure, morphosyntactic feature specifications, and so forth). The role of the phonology is then simply to compute an output (surface) phonological representation of this object. Viewed from the function perspective, this model can be depicted as ϕ(µ( · )) or more accurately as ϕ(µn ( . . . (µ1 ( · )))). In such a model, there are in principle no limitations on the ways in which the constraints or rules of phonology might make reference to, and thus be sensitive to, morphological factors. We thus expect extensive influence of morphology on phonology (M→P), but there should be little or no influence of phonology on morphology (P→M), though morphology might perhaps be sensitive to the lexical phonological properties (that is, ones not derived by ϕ) of the stems and exponents that it manipulates. At the other extreme, we can imagine that all phonological and morphological structure is assembled and computed in parallel, such that each is fully capable of influencing the other. If implemented in terms of a constraint-based architecture such as Optimality Theory (McCarthy 2002, 2007; Prince and Smolensky 2004), this is in essence a non-modular model, in which constraints that reference either morphological or phonological content (or both at once) are fully interspersed and each type of constraint can outrank the other. It is thus not only possible for consideration of morphological factors to affect phonological patterning (M ≫ P) but likewise for aspects of morphology (e.g. exponence, affix order) to be affected by constraints on phonological well-formedness (P ≫ M). The full-parallellism model differs from the all-morphology-precedes-phonology model primarily in that there is a much less clear distinction between the functions ϕ and µ, with much of what is normally attributed to the latter being instead folded into the former (see Embick 2010 and Trommer and Zimmerman, this volume). In practice, existing approaches to modelling the interface and interaction between phonology and morphology fall somewhere between these two straw-man extremes. Each of the two is at once too permissive and unconstrained and too limited in its abilities to adequately capture the richness of attested phenomena. In what follows, we first (Section 8.2.1) consider a class of models that incorporate morphology– phonology interleaving, such that phonology is applied and re-applied in a cyclic fashion at intermediate nodes in the constituent structure. In models of this type, construction-specific phonology is accounted for by allowing successive cycles to call on distinct phonology functions, or co-phonologies (levels, strata). The remaining subsections critically examine various alternative proposals for non-interleaving and monostratal accounts of some of the same empirical phenomena. Section 8.2.2 discusses the use of constraint indexation, as well as reference to internal morphosyntactic constituent boundaries, as a means of capturing construction-specific phonological patterns. Section 8.2.3 considers output–output correspondence as an alternative to interleaving in accounting for so-called cyclicity effects (see also Section 8.3.1 on the related notion of paradigm uniformity).

164

gunnar ólafur hansson

8.2.1 Interleaving: Cycles, Strata, Co-phonologies One influential model of morphology–phonology interaction, which had its roots in Chomsky and Halle’s (1968) cyclic account of English stress but was more fully developed in the Lexical Phonology model (Kiparsky 1982a, 1982b; Kaisse and Shaw 1985; Mohanan 1986; Kaisse and McMahon 2011) and its constraint-based descendants, takes the phonological interpretation of complex morphological structures as involving an interleaving, or layering, of phonology and morphology. There is thus not merely a single call to the phonology function ϕ at the very end of the derivation of a morphologically complex form. Rather, the derivation may involve one or more internal passes through phonology as well: ϕ(µn ( . . . (ϕ( . . . (µ1 (·)))))). In other words, application of some phonological processes can precede certain aspects of morphology, creating the potential for various interactions and interference—between phonology and morphology, and between inner and outer applications of phonology—which would otherwise be inconceivable. In a simplified view of morphology as involving only morpheme concatenation, interleaving can be illustrated schematically as in (1), where phonological interpretation takes place at each non-terminal node in a constituent-structure tree. Here {a, b, c, d} represent individual affix morphemes, of which {a, b} are cyclic, each defining a domain for phonological interpretation, whereas {c, d} are non-cyclic. At each node, M stands for its morphological constituent structure (category labels on constituents are left out for simplicity) and 8 for its phonological content; a is the string-concatenation operation. (1)

8 = ϕk (ϕj (b a ϕi (X a a)) a c a d) M=[[b[Xa]]c d]

8 = ϕj (b a ϕi (X a a))

c

d

M=[b[Xa]]

b

8 = ϕi (X a a) M=[Xa]

X

a

The innermost constituent, [X a], has its phonological properties computed by applying phonology (ϕi ) to the concatenation of the underlying phonological representations of the root (X) and the innermost suffix (a): ϕi (X a a). This derived

phonology

165

representation in turn acts as one of the two concatenated terms at the next level up, where the prefix (b) is attached and phonology is once again applied. The output of the phonological computation at this node in the tree is therefore ϕj (b a ϕi (X a a)). Finally, this representation is concatenated with those of the two outermost suffixes (c, d) and phonology is applied yet again. The output phonological representation of the word form as a whole is therefore defined as ϕk (ϕj (b a ϕi (X a a))), with three separate, and nested, calls to phonology. In (1), the successive applications of phonology have been labelled ϕi , ϕj , ϕk . If all of these are one and the same function, ϕi = ϕj = ϕk = ϕL , then what we are seeing is pure cyclicity: the cyclic application of a single phonology function (constraint ranking, or set of rules) as dictated by the morphological structure. In practice, however, interleaving models generally presuppose the co-existence of more than one phonological sub-grammar. Models differ with respect to whether any limit is placed on how many such co-existing phonologies, ‘cophonologies’, the grammar can contain or on how these may differ from each other, and whether the cophonologies are assumed to be strictly ordered relative to one another. Stratal Optimality Theory (Kiparsky 2000, 2008a; Bermúdez-Otero 2003, 2012), the constraint-based implementation of Lexical Phonology (Kiparsky 1982a, b; Mohanan 1986), posits exactly three distinct levels, or strata, of phonology: Stem, Word, and Phrase. These are strictly ordered, and any phonological derivation must involve at least one call to each of the three: ϕL ( · ) = ϕPhrase ( . . . (ϕWord ( . . . (ϕStem ( . . . ( · )))))). For example, Kiparsky (2000) draws upon the Stem vs. Word distinction to account for the famously opaque interaction of syncope, stress, epenthesis, and morphological structure in Levantine Arabic (Brame 1974; Kenstowicz 1983). The basic pattern of stress assignment is that a penultimate syllable will attract stress if it is heavy (e.g. CVC or CVː) but otherwise stress falls on the antepenultimate syllable. Syncope is sensitive to stress placement, targeting short /i/ in unstressed open (and hence light) syllables. Finally, complex syllable onsets or codas are not permitted, and epenthesis of [i] breaks up otherwise unsyllabifiable consonant clusters of three or more consonants. The puzzle, which is illustrated by the minimal triplet in (2), is twofold. First, assignment of stress to the penultimate syllable triggers syncope of unstressed /i/ before a subject marker like 1.pl /-na/ (2a), but mysteriously fails to do so before the corresponding object or possessive markers (2b). Second, epenthesis into a morphologically derived triconsonantal cluster, such as when 1.pl /-na/ is attached to the noun /fihm/ (2c), fails to trigger stress shift to the penultimate epenthetic vowel, despite the fact that it is in a heavy syllable. (2)

Strata and cycles in Palestinian Arabic a. [ˈfhim.na] /fihim-na/ ‘we understood’ b. [fi.ˈhim.na] /fihim-na/ ‘he understood us’ c. [ˈfi.him.na] /fihm-na/ ‘our understanding’

(Kiparsky 2000) [ [ [ [ fihim ]St na ]St ]Wd ]Phr [ [ [ [ fihim ]St ]Wd na ]Wd ]Phr [ [ [ [ fihm ]St ]Wd na ]Wd ]Phr

166

gunnar ólafur hansson

Kiparsky’s (2000) stratal analysis relies on the insight, supported by a variety of other sound patterns as well, that object/possessive markers in Levantine Arabic are attached at the word level whereas subject markers are stem-level (see, e.g., Kenstowicz 1983), as shown by the bracketing in the rightmost column of (2). Given certain differences in constraint ranking between the three strata, the facts fall out correctly. The failure of epenthesis to affect stress in (2c) is explained by the fact that epenthesis is a Phrase-level process (cf. /fihm l-walad/ → [ˈfih.m il.-ˈwa.lad] ‘the boy’s understanding’, not *[ˈfi.him. il.-ˈwa.lad]), whereas weight-sensitive stress assignment is (re)computed solely at the Stem and Word levels. That is, Phrase-level phonology simply preserves the input (Word-level) stress pattern faithfully. The failure of syncope to apply in (2b) arises from the fact that syncope, a Word-level process, is blocked from deleting vowels that are stressed in the input (i.e. ones which have been assigned stress in the Stem-level ´ phonology), due to a high-ranked faithfulness constraint Max-V-IO. In terms of phonology, the derivational history of the three forms in (2) is then as shown in (3). Note that in each case, the affixation of /-na/ triggers a new cycle through the phonology of the relevant (Stem or Word) level. Note also that the rootand-template morphology of Arabic, which combines the triconsonantal root /fhm/ with the appropriate CVCVC or CVCC template and vocalism, is likewise Stem-level but has been omitted here for simplicity. (3)

‘we understood’

‘he understood us’

‘our understanding’

Stem: Stem (2nd cycle): Word: Word (2nd cycle): Phrase:

ˈfi.him fi.ˈhim.na ˈfhim.na (n/a)

ˈfi.him (n/a) ˈfi.him fi.ˈhim.na

ˈfihm (n/a) ˈfihm ˈfihm.na ˈfi.him.na

Output:

[ˈfhim.na]

[fi.ˈhim.na]

[ˈfi.him.na]

An alternative to stratal Optimality Theory, cophonology theory (Orgun 1996, 1999; Inkelas 1998, to appear; Inkelas and Orgun 1998; Anttila 2002; Inkelas and Zoll 2005, 2007), abandons several of its restrictive assumptions and generalizes the stratum notion by extending it to cover all construction-specific phonological patterns. In this model, each morphological construction is associated with some phonological function (constraint ranking), a cophonology, and there is in principle no limit on how many cophonologies can co-exist within the grammar, nor how many individual morphological constructions ‘subscribe’ to any given cophonology. Importantly, cophonologies are not assumed to be subject to any crucial ordering. For some pair of cophonologies ϕi ̸= ϕj , the fact that the order ϕj ( . . . (ϕi ( . . . ))) occurs does not automatically preclude the possibility of the reverse order ϕi ( . . . (ϕj ( . . . ))), as long as the morphological structure so dictates. Furthermore, because cophonologies are tied directly to morphological constructions, their application is economical (in the sense

phonology

167

of Inkelas and Orgun 1995): unlike in stratal Optimality Theory, there is no requirement that a given word form must be subjected to each cophonology (level, stratum) in the language. Finally, there is no stipulated (theory-internal) limit on the extent to which cophonologies, or the levels of stratal Optimality Theory, may in principle diverge. Such divergence will in any event be constrained in the diachronic domain of language transmission and change (on this point see also Bermúdez-Otero and McMahon 2006; Pater 2010). Anttila (2002) approaches the question of shared vs. non-shared properties among cophonologies by modelling the phonological grammar as a whole as an inheritance hierarchy. Each node in the hierarchy defines a particular constraint ranking (cophonology), inheriting and combining ranking information from all higher nodes, such that the default or unmarked state of affairs is for cophonologies to be quite similar. In Anttila’s account of the declension of Finnish /a/-final nominals, the node to which a given class of nominal belongs determines the choice between mutation (/a/ → [o]) or deletion (/a/ → ∅) before plural /-i-/, the degree to which that choice is sensitive to syllable count (even- vs. odd-numbered stems) and/or the quality of the penultimate stem vowel, and the extent to which the choice is variable.

8.2.2 Parallellism: Constraint Indexation, Boundary Reference A different line of scholarship eschews morphology–phonology interleaving and cophonologies in favour of a monostratal model (i.e. one employing a single constraint ranking), in which apparent interleaving effects are instead accounted for by other means. In the Optimality Theory tradition, a popular alternative to strata or cophonologies is to allow for certain constraints to be indexed to particular morphological constituents or lexical categories, or even to individual morphemes, lexemes or lexical strata (McCarthy and Prince 1995; Itô and Mester 1999; Pater 2000, 2010; Alderete 2001; Smith 2011). In the most common instantiation of this approach, only correspondence constraints can be indexed in this manner, in particular those regulating faithfulness (input–output correspondence). As an illustrative example, consider the two different strategies for hiatus resolution that are employed in Turkish in different morphological contexts (Inkelas and Zoll 2007). The general pattern is epenthesis of the glide [j] between the stem-final and suffix-initial vowels, as illustrated in (4). However, before the progressive /-Ijor/ suffix, the stem-final vowel instead deletes, as seen in (5).2 2

The archiphonemic symbols /I A/ denote a vowel that is underlyingly specified as [+high] or [−high], respectively, but whose value for backness (and in the case of /I/, rounding) is determined by harmony to the vowel in the preceding syllable.

168

gunnar ólafur hansson

(4) Turkish: glide epenthesis as default hiatus resolution strategy a. [adam-a] [kapɯ-ja]

/adam-A/ /kapɯ-A/

man-dat door-dat

‘to a/the man’ ‘to a/the door’

b. [ev-i] [oda-jɯ]

/ev-I/ /oda-I/

house-def.acc room-def.acc

‘the house (obj.)’ ‘the room (obj.)’

c. [koʃ-aʤak] [anla-jaʤak]

/koʃ-AʤAk/ /anla-AʤAk/

run-fut.ptcp understand-fut.ptcp

‘he will run’ ‘he will understand’

d. [al-ɯnʤa] [dene-jinʤe]

/al-InʤA/ /dene-InʤA/

take-cvb try-cvb

‘upon taking’ ‘upon trying’

(5) Turkish: hiatus resolution by vowel deletion in progressives a. [jap-ɯjor] [koʃ-ujor]

/jap-Ijor/ /koʃ-Ijor/

make-prog run-prog

‘he is making’ ‘he is running’

b. [anl-ɯjor] [ød-yjor] [koʃ-m-ujor]

/anla-Ijor/ /øde-Ijor/ /koʃ-mA-Ijor/

understand-prog pay-prog run-neg-prog

‘he is understanding’ ‘he is paying’ ‘he is not running’

As Inkelas and Zoll (2007) note, these facts could easily be captured in an indexedconstraint analysis, by assuming a constraint ranking such as that in (6). Here NoHiatus stands for whatever markedness constraint is responsible for prohibiting hiatus in Turkish. (This may well be the syllable structure constraint Onset, which bans onsetless syllables.) As input–output correspondence constraints, Max-IO penalizes deletion whereas Dep-IO bans epenthesis (McCarthy and Prince 1995). (6) { NoHiatus, Dep-IOprog } ≫ Max-IO ≫ Dep-IO The indexed constraint Dep-IOprog is relevant only for segments contained within a Stem + /-Ijor/ constituent. That constraint thus has the effect of preventing [j]-epenthesis in progressive forms specifically, such that the alternative repair strategy of vowel deletion emerges as optimal (7). In non-progressive contexts, where the indexed constraint is irrelevant (and hence vacuously satisfied), the default repair of [j]-epenthesis occurs (8). (7) /anla-Ijor/ → [anl-ɯjor] ‘he is understanding’ /anla-Ijor/ a. b. c.

an.la.ɯ.jor an.la.jɯ.jor an.lɯ.jor

NoHiatus *!

Dep-IOprog *!

Max-IO

*

Dep-IO *

phonology

169

(8) /anla-InʤA/ → [anla-jɯnʤa] ‘upon understanding’ /anla-InʤA/ a. b. c.

an.la.ɯn.ʤa an.la.jɯn.ʤa an.lɯn.ʤa

NoHiatus

Dep-IOprog

Max-IO

*! *!

Dep-IO *

In this way, constraint indexation is able to mimic what in a cophonology analysis would be a matter of different ranking relations among the same set of three constraints: { NoHiatus, Dep-IO } ≫ Max-IO in the progressive-construction cophonology, and { NoHiatus, Max-IO } ≫ Dep-IO otherwise. The two approaches are not merely notational equivalents of one another, however. In particular, cophonology theory is embedded in an interleaving model of the morphology–phonology interface (cf. (1) above). Constituent structure and nesting relations are thus predicted to dictate the interaction and trade-off between the cophonologies (rankings) that are called upon in the derivation of a morphologically complex word form. Consider two morphological operations µj and µk (e.g. prefixation and suffixation, or zero-derivation and affixation). If the morphology of the language allows either to be nested within the other, such that the constructions µj (µk (·)) and µk (µj (·)) both occur, resulting in the same linear string of morphemes, cophonology theory predicts the phonological outputs for the two to be different. This is due to the different nesting relations between the relevant cophonologies: in the former case, ϕj applies to the output of ϕk , and vice versa in the latter. In either case, the ‘outer’ cophonology can be sensitive to properties induced by the ‘inner’ cophonology, or potentially overwrite them. One example of this sort of interaction is found in stress patterns associated with derived place names in Turkish (Inkelas and Orgun 1998, 2003; Inkelas 1999). In Turkish, the default word stress falls on the word-final syllable. However, a productive zero-derivation process, by which any word can be converted into a place name, is associated with a special cophonology in which the word (if not already stressed) receives stress on the antepenultimate syllable if it is heavy (CVC) and the penultimate syllable is light (CV), or else on the penultimate syllable. This results in pairs of place names and the words from which they are derived which differ only in the location of stress, such as [ba.ˈka.ʤak] Bakacak (place name) vs. [ba.ka.ˈʤak] ← /bak-AʤAk/ (look-fut) ‘s/he will look’, or [ˈsir.ke.ʤi] Sirkeci (place name) vs. [sir.ke.ˈʤi] ← /sirkeˈʤI/ (vinegar-agt) ‘vinegar seller’. In addition, Turkish also has a large set of suffixes that subscribe to a different, pre-stressing cophonology, which assigns stress to the syllable immediately preceding the suffix. This includes /-ʤA/ mitigative ‘kind of; in the manner of; (in the) language of ’, e.g. [syt.ˈly.ʤe] ← /syt-lI-ʤA/ (milk-assoc-mit) ‘kind of milky’. The crucial evidence comes from cases where a pre-stressing suffix is either attached to a derived place name, or is contained within it, as the two cophonologies may then place conflicting demands on stress assignment. For example, the place name Çamlıca is derived from a form with the same internal structure as [syt.ˈly.ʤe] just mentioned (/ʧam-lI-ʤA/, pine-assoc-mit); it retains that same stress

170

gunnar ólafur hansson

pattern, [ʧam.ˈlɯ.ʤa], rather than adopting the expected place-name stress pattern (*[ˈʧam.lɯ.ʤa]).3 Conversely, if the pre-stressing suffix /-ʤA/ is attached to a place name like Kavaklι [ka.ˈvak.lɯ], a zero-derivation from /kavak-lI/ (poplar-assoc) ‘with poplar’, it is the place-name stress pattern that prevails, [ka.ˈvak.lɯ.ʤa] ‘(in the) language/dialect of Kavaklı’, not the pre-stressing one (*[ka.vak.ˈlɯ.ʤa], which would have been analogous to [ʧam.ˈlɯ.ʤa]). The observed outcome is thus in all cases consistent with the ‘innermost wins’ principle that appears to guide stress assignment throughout the Turkish lexicon (see Inkelas and Orgun 2003). This interaction pattern falls out straightforwardly in the cophonology model because of how that model takes phonology to be interleaved with, and therefore inevitably reflective of, constituent structure. All that is required is to assume that stress assignment in Turkish is structure-building, applying only to forms that are not already stressed (e.g. due to containing a lexically stressed root or suffix; Inkelas and Orgun 2003: 141–2). As a consequence, whichever stress-assigning cophonology happens to be innermost determines stress placement (in the absence of lexical stress); any outer layers of phonology are unable to alter this in any way. As Inkelas (2011: pp. 74–75) points out, the constraint indexation model does not fare as well. Since it assumes a single constraint ranking, the constraints crucially associated with the pre-stressing pattern must either outrank or be outranked by the constraints that drive the place-name stress pattern (as well as those that define the default, word-final stress pattern). The erroneous prediction is thus that we should see either consistent pre-stressing ([ʧam.ˈlɯ.ʤa], *[ka.vak.ˈlɯ.ʤa]) or consistent enforcement of the place-name stress pattern (*[ˈʧam.lɨ.ʤa], [ka.ˈvak.lɯ.ʤa]), regardless of scope relations between the zero-derivation and /-ʤA/ suffixation. One argument that is sometimes adduced in support of the constraint indexation approach over interleaving approaches (cophonologies, strata) is that it is supposedly more restrictive, in that construction-specific phonology is predicted to exhibit what Alderete (2001) refers to as grammar dependence (see also Benua 1997; Itô and Mester 1999). Indexed correspondence constraints are interdigitated within the overall ranking of markedness and (other) correspondence constraints in the language, and this is claimed to restrict the degree of language-internal phonological variation that morphological constructions can exhibit; in particular, markedness reversals are claimed to be impossible. However, Inkelas and Zoll (2007) show how this gain in restrictiveness is largely illusory, and in any case crucially depends on the stipulation that indexation is limited to correspondence constraints. Much recent work, by contrast, has argued that indexation of markedness constraints must also be allowed for (Pater 2000, 2007, 2010; Ota 2004; Flack 2007). A different way in which sensitivity to morphological structure can be captured in a monostratal model is by reference to constituent edges or boundaries. Explicit use

3 Thanks to Orhan Orgun for providing native-speaker judgements on the relevant place name forms cited in this paragraph.

phonology

171

of juncture or boundary symbols was commonplace in classical generative phonology (Chomsky and Halle 1968) as well as earlier structuralist analyses, and boundary reference remains a relevant mechanism for morphology–phonology interaction in current constraint-based frameworks. In particular, alignment constraints (McCarthy and Prince 1993b) can reference the edges of morphological constituents of various types—either directly or else indirectly via such ‘morpho-prosodic’ constituents as the prosodic word, stem, or root (PWord, PStem, PRoot; Nespor and Vogel 1986; Inkelas 1989, 1993; Downing 1998, 1999, 2006; Raffelsiefen 2005; cf. also Selkirk 1986). For example, in an analysis of hiatus resolution strategies in Shona, Mudzingwa (2010) proposes that a noun class prefix combines with a following (nonderived) nominal stem to form a PStem constituent. However, derived nominal stems (e.g. deverbal agentives and infinitives) also constitute their own PStem domain, whereas non-derived stems do not, resulting in a recursive domain structure in the former case.4 A noun class prefix will therefore be separated from the following nominal stem by a PStem boundary only if the latter is derived (9b) but not otherwise (9a). (9) Hiatus resolution and constituent structure in Shona a.

(Mudzingwa 2010)

Non-derived nominal stems: [PWord [PStem Prefix Stem ] ] [mw ì.sé] /mù-ìsé/ cl3.sg-tail ‘tail’ [ʋó.sé] /ʋà-ósé/ cl2.pl-all ‘all [the boys]’

b. Derived (deverbal) stems: [PWord [PStem Prefix [PStem Stem ] ] ] [mù.jí.m͡bí] /mù-ím͡b-í/ cl1.sg-sing-nmlz ‘singer’ [ʋà.wó.ŋ͡ɡó.ró.rí] /ʋà-óŋ͡ɡórór-í/ cl2.pl-spy-nmlz ‘spies’ This difference in morphological (and morpho-prosodic) structure correlates with a difference in how hiatus is repaired at the prefix–stem juncture. PStem-internal hiatus as in (9a) is resolved by elision, glide formation, or consonant-vowel fusion (/Cu/ → [Cw ]), depending on the segmental context. Hiatus across a PStem boundary as in (9b), however, is consistently resolved by epenthesis (of [j], [w], [ʔ], or [ɦ], again depending on segmental context). Mudzingwa (2010) analyses this as being due to an alignment constraint Align(PStem, L, σ, L) which demands that the left edge of a PStem coincide with the left edge of a syllable.

8.2.3 Cross-derivational Interactions In the approaches discussed in the previous sections, the phonological component accesses information about morphological structure by means of making reference to morphological constituents as such, or to their edges, either directly or indirectly 4 The same type of recursion arises in noun-class prefix stacking (so-called pre-prefixation) as well. For example, [má.zì.wú.mé] /má-zì-ú-mé/ (cl6.pl-cl21.aug-cl14-fruit) ‘big ume fruits’ has the domain structure [PWord [PStem má [PStem zì [PStem ú - mé ] ] ] ].

172

gunnar ólafur hansson

(via a ‘morpho-prosodic’ domain structure). This mode of morphology–phonology interaction can thus only refer to structural properties that are inherent in the word form in question; there is no explicit reference to other morphologically related words or word forms, such as other forms within the same inflectional paradigm. In the context of Optimality Theory, alternative approaches have been developed which instead (or in addition) allow direct reference to another output word form that is a member of the same morphological paradigm, or even to the entire set of output forms that make up the inflectional paradigm of the word in question. The former of these proposals relies on a formal device usually referred to as transderivational, or output–output, correspondence (Benua 1995, 1997; Crosswhite 1998; Kager 1999: ch. 6: see also Burzio 1994, 1998, 2002a and the notion of base identity in Kenstowicz 1996). Constraints that demand correspondence between a given output form and a designated ‘base’ form in the same morphological paradigm can have the effect of either extending (overapplying) the effects of a phonological process that would otherwise not have been applicable, or conversely suppressing (underapplying) a process that ought to have affected the word form in question. This is illustrated in (10), which depicts the Levantine Arabic case discussed in (2)–(3) above. The vertical arrows denote the dimension of input-to-output correspondence (faithfulness), the horizontal arrow the added dimension of base-toaffixed-form (output-to-output) correspondence. The difference between the forms [ˈfhim.na] ‘we understood’, in which syncope is enforced, and [fi.ˈhim.na] ‘he understood us’, where it fails to apply, is that the second form is also subject to correspondence to the independently-occurring unsuffixed form [ˈfi.him] ‘he understood’. The forms ‘he understood’ and ‘we understood’, by contrast, do not constitute a base vs. affixed form pair in this sense, due to the fact that the morphosyntactic feature content of the former is not a proper subset of the latter (Kager 1999: ch. 6; though see Kiparsky 2000 on the problems with this criterion). (10)

‘we understood’

‘he understood’

‘he understood us’

/fihim-na/

/fihim/

/fihim-na/

[ˈfi.him]

B–A corr.

−−−−−−−−→

I–O corr.

−−−−−−−−→

I–O corr.

−−−−−−−−→

I–O corr.

−−−−−−−−→

[ˈfhim.na]

[fi.ˈhim.na]

Here the base [ˈfi.him] plays a similar role in the derivation of the affixed form [fi.ˈhim.na] ‘he understood us’ as it did in the stratal (cyclic) account described in example (3). In the stratal account, [ˈfi.him] constituted an intermediate representation in the overall derivation: the output of the Stem stratum. This meant that the initial-syllable [i] was already stressed in the input to the Word stratum, and protected

phonology

173

against syncope in that stratum by a high-ranked input–output correspondence con´ straint Max-V-IO. The monostratal account portrayed in (10) utilizes that very same ´ type of constraint, but defined instead over the output–output dimension: Max-V-BA (called HeadMax-BA in Kager 1999: 283). One limitation of the transderivational correspondence approach is that it crucially requires that the source of influence (the ‘base’) occur independently as a fully fledged output form. This can create challenges in situations where the putative base is not a morphologically simplex form, as is often the case for inflectional morphology. A related problem is how the criteria for the unique identification of bases should be formulated, especially in inflectional paradigms. A related line of thinking avoids the problem of base identification by appealing to a broader and more symmetric concept of paradigm uniformity, or anti-allomorphy (Kenstowicz 1996; Steriade 2000; Burzio 2002a; McCarthy 2005). Since such proposals often explicitly invoke the notion of inflectional paradigm, they are discussed in Section 8.3.1.

8.3 The Phonology of Paradigms

.............................................................................................................................................................................

8.3.1 Levelling and Paradigm Uniformity The most obvious way in which paradigmatic relations can have repercussions for the phonological structure of word forms is in the phenomenon of levelling, whereby an otherwise expected phonological alternation in the surface shape of some item (morpheme, stem, lexeme) is either suppressed or generalized throughout the inflectional paradigm. The Palestinian Arabic example in the preceding section can be seen as an instance of levelling, in that the syncope alternation that is evident in the subject–agreement paradigm (e.g. [ˈfihim] ‘he understood’, [ˈfhim-na] ‘we understood’, [ˈfihm-u] ‘they understood’) is suppressed in the sub-paradigm defined by object-agreement suffixation ([fiˈhim-na] ‘he understood us’, etc.).5 For the present purposes, the term ‘(paradigm) levelling’ will be used as a purely descriptive label, referring to any synchronic state of affairs that fits the above definition. As we will see toward the end of this section, it is far from certain that levelling exists as a distinct type of process in its own right. Various works in the theoretical phonology literature propose that the synchronic source of at least some levelling phenomena is a demand for paradigm uniformity: the uniform exponence of some morphological entity (e.g. the stem) throughout the paradigm (Steriade 2000; Kenstowicz 1996; Burzio 2002a; McCarthy 2005), rather than 5

This is a slight oversimplification, in that it is only with respect to the vowel in the stem-initial syllable that syncope alternations are suppressed (the vowel that is stressed in the unsuffixed form [ˈfihim] ‘he understood’). Syncope applies as expected in the second syllable when its conditions are met: /fihim-u/ → [ˈfih.mu] ‘he understood them’.

174

gunnar ólafur hansson

identity to some designated base form. The optimal paradigms model of McCarthy (2005), for example, takes it as a given that inflectional paradigms do not have bases (for counterarguments, see Hall and Scott 2007; Albright 2008, 2010; Bobaljik 2008). Instead of evaluating the well-formedness of individual word forms in the paradigm against the benchmark of some designated base form within that same paradigm (as in Section 8.2.3), the overall phonological well-formedness of the entire paradigm is assessed all at once. The candidate outputs under evaluation by the phonology are thus not single word forms but whole sets of forms that together make up an inflectional paradigm. Paradigm uniformity is enforced by correspondence constraints that compare the stem portions of every pair of output forms that are contained within the (candidate) paradigm. Consider, for example, a hypothetical language in which high vowels are obligatorily lowered before uvulars (*HiV/__Uvu ≫ Ident[±high]-IO), and whose nominal inflection includes a plural suffix /-q/. In such a language, a high-ranked paradigm uniformity constraint Ident[±high]-OP will have the effect shown in (11). The process of pre-uvular lowering overapplies, in that its effects become extended even to the unsuffixed form where no uvular is present. (11) Levelling (by overapplication) due to paradigm uniformity /sani/ + {∅, q, . . . } a. b. c.

⟨sani, saniq, . . . ⟩ ⟨sani, saneq, . . . ⟩ ⟨sane, saneq, . . . ⟩

Ident[±high]-OP *HiV/__Uvu Ident[±high]-IO *!

*!

* **

As McCarthy (2005) points out, the inherently asymmetric nature of paradigm uniformity makes specific predictions regarding patterns of levelling in inflectional (as opposed to derivational) paradigms, predictions which would not be consistent with either a base-identity approach (Section 8.2.3) or a cyclic/stratal approach (Section 8.2.1). For example, the forms that act as attractors should be ones that are phonologically less marked. Furthermore, ‘majority rule’ effects are expected to be possible, in which the sheer number of forms exhibiting each alternant (each stem allomorph, say) will determine the direction of levelling. Finally, inflectional paradigms should never exhibit levelling that is manifested through under- rather than overapplication of a phonological process. It should be noted, however, that these predictions of the paradigm uniformity hypothesis pertain to levelling in the sense of systematic, regular patterns that can reasonably be ascribed to the synchronic grammar of the language in question. The question of how (or whether) these considerations apply to levelling as a diachronic process will be taken up at the end of this section. The paradigm uniformity notion has potential applications that may at first glance seem counterintuitive. McCarthy (1998, 2005) uses it as a means of accounting for templatic or canonical shape restrictions on stems or roots. For example, McCarthy (2005) calls attention to the fact that Classical Arabic verb stems can only end in /. . . CVC/, whereas noun stems are less restricted and can also end in /. . . CVːC/ or /. . . CVCC/.

phonology

175

By appealing to paradigm uniformity constraints, he explains this as a mere by-product of the (accidental) fact that the inflectional suffixes of nouns in Classical Arabic all happen to be vowel-initial, whereas the inflectional morphology of verbs includes consonant-initial suffixes as well. Crucially, the phonology of Classical Arabic permits neither tri-consonantal clusters nor superheavy syllables, and stem–suffix sequences like /. . . CVCC–CV/ or /. . . CVːC–CV/ would thus require repair by vowel epenthesis and vowel shortening, respectively. A high-ranked demand for paradigm uniformity would have the effect of extending (overapplying) this repair throughout the inflectional paradigm. As a result, even if an underlying verb stem like /faʕaːl/ were to exist, its surface manifestation would be indistinguishable from that of its /CVCVC/ counterpart /faʕal/, as illustrated in (12). The input /faʕaːl/ would thus not be recoverable from the point of view of the language learner (an instance of what Prince and Smolensky 2004 dub ‘Stampean occultation’). The complete absence of consonant-initial suffixes in the nominal paradigm, by contrast, means that the threat of vowel-length alternations does not materialize there, and a stem of the underlying shape /CVCVːC/ is thus free to surface intact, and is clearly distinct from a /CVCVC/ stem, as seen in (13). (12) Inflectional paradigm of hypothetical verb stem /faʕaːl/ /faʕaːl/ + {a, tu, . . . } a. b. c.

⟨faʕaːla, faʕaːltu, . . . ⟩ ⟨faʕaːla, faʕaltu, . . . ⟩ ⟨faʕala, faʕaltu, . . . ⟩

Ident-µ-OP *!

*σµµµ *!

Ident-µ-IO * **

(13) Inflectional paradigm of noun stem /kitaːb/ ‘book’ /kitaːb/ + {u, i, . . . } a. b. c.

⟨kitaːbu, kitaːbi, . . . ⟩ ⟨kitaːbu, kitabi, . . . ⟩ ⟨kitabu, kitabi, . . . ⟩

Ident-µ-OP *!

*σµµµ

Ident-µ-IO * *!*

However, as Bobaljik (2008) points out, the absolute neutralization of /CVCVC/ and (hypothetical) /CVCVːC/ verb stems at the level of the inflectional paradigm, seen in (12), is insufficient as an explanation for the categorical absence of /CVCVːC/ verb stems. This is because the underlying vowel length contrast should still manifest itself overtly in other, non-inflectional morphology, such as nominalizations. Such evidence would provide the necessary clues for learners in order to internalize the appropriate /CVCVːC/ or /CVCVC/ underlying representation for each verb stem. To cite a parallel from English, derivational formations like [dæm.n-eɪ.ʃən] provide evidence for the verb damn having an underlying representation /dæmn/, with final /n/, despite the complete neutralization toward [n]-less [dæm] throughout the inflectional paradigm ([dæm], [dæm-d], [dæm-ɪŋ], etc.)—which is similarly due to levelling by overapplication, in this case of coda cluster simplification.

176

gunnar ólafur hansson

What unites all accounts of levelling phenomena that attribute them either to output–output correspondence (base identity) or some notion of paradigm uniformity is one fundamental assumption: that there exists a universal desideratum or preference for identity of exponence, and thus for the absence of morphophonological alternations that disrupt such identity. As was noted earlier, proposals for paradigm uniformity constraints are intended to account for cases in which there appears to be evidence for levelling as a systematic, synchronic pattern that can thus be attributed to the grammar of the language in question (as a model of speakers’ implicit knowledge). What the relationship is between this notion and that of levelling as a diachronic process (often sporadic) is a question that is largely left unaddressed in the phonological literature advocating these approaches. While the idea of uniform exponence certainly has deep roots in the historical linguistics tradition (see e.g. Hock 1991: 168)—and is indeed implicit in the traditional term ‘paradigm levelling’—the notion of levelling as a distinct type of language change has increasingly come under challenge in that field in recent years. Evidence now seems to suggest, rather, that all putative cases of paradigm levelling as a diachronic process may instead be instances of analogical extension, in that levelling appears to crucially depend on the prior existence of a pattern of non-alternation that can serve as an analogical model (Albright 2005; Garrett 2008; see also Hill 2007). Albright (2005) argues against paradigm uniformity explanations of the historical levelling of Latin rhotacism alternations ([s] ∼ [r]), instead advocating an account that is based on a formally explicit model of morphological learning. In effect, this attributes paradigm levelling to imperfect learning, which has long been the standard account for analogical change in general (Kiparsky 1965, 1968).

8.3.2 Paradigm Anti-homophony Effects As noted in the previous section, levelling and paradigm uniformity have the effect of reducing phonological differences among the set of word forms that make up an inflectional paradigm, by eliminating alternations within the stem portion that these forms share. Such anti-allomorphy effects can in principle involve either the suppression of the operation of a phonological process or an extension of its application into contexts which do not fit its conditioning environment. Conversely, departures from the normal application of a phonological process sometimes appear to serve the goal of maintaining or creating a difference in phonological form between morphosyntactically distinct forms within the paradigm. Such phenomena are usually referred to as anti-homophony effects, or intra-paradigmatic homophony avoidance (Crosswhite 1999; Blevins 2005; Gessner and Hansson 2005; Ichimura 2006; Blevins and Wedel 2009; Mondon 2009), though perhaps ‘syncretism avoidance’ would be a more apt term. Just as levelling has been attributed to a general principle (or constraint) of paradigm uniformity, so anti-homophony effects have been similarly explained by

phonology

177

invoking a principle of paradigmatic contrast and its maintenance (Kenstowicz 2005; Rebrus and Törkenczy 2005; see also Alderete et al. 1999; Kurisu 2001 for related notions). As an example, consider Kenstowicz’s (2005) discussion of various phonological irregularities that characterize 3sg.f perfective forms of verbs in many modern Arabic dialects, specifically when these are followed by an object suffix. In the Damascus dialect of Syrian Arabic, for example, such forms exhibit an aberrant stress pattern in precisely those cases where regular stress, and concomitant syncope of unstressed schwa in open syllables, would have resulted in homophony with the corresponding 1sg and 2sg.m forms. (The latter two are themselves homophonous/syncretic, a point to which we will return later in this section.) This is illustrated in (14). The aberrant 3sg.f forms, and the 1sg/2sg.m forms with which they ought to be homophonous, were the phonology of stress and syncope to operate as expected, are shown in boldface. The boundary between the stem and subject-agreement suffix is indicated by a hyphen, that between the subject and object suffixes by #. The regular stress pattern of Syrian Arabic is that stress is limited to a three-syllable window at the right edge of the word; within this window, stress falls on the rightmost heavy syllable (a wordfinal consonant counts as extrametrical and does not render the final syllable heavy), otherwise on the antepenultimate syllable. (14) Damascus Arabic: anti-homophony effects in perfective paradigms (Kenstowicz 2005) Object

‘he taught’

‘she taught’

‘I taught / you-sg.m taught’

(none)

ˈʕal.lam

ˈʕal.la.m-ət

ʕal.ˈlam-t

1sg 2sg.m 2sg.f 3sg.m 3sg.f 1pl 2pl 3pl

ʕal.ˈlam.#ni ˈʕal.la.m#ak ˈʕal.la.m#ək ˈʕal.la.m#o ʕal.ˈlam.#(h)a ˈʕal.ˈlam.#na ʕal.ˈlam.#kom ʕal.ˈlam.#(h)om

ʕal.la.ˈm-ət.#ni ʕal.la.ˈm-ə.t#ak ʕal.la.ˈm-ə.t#ək ʕal.la.ˈm-ə.t#o ʕal.la.ˈm-ət.#(h)a ʕal.la.ˈm-ət.#na ʕal.la.ˈm-ət.#kom ʕal.la.ˈm-ət.#(h)om

ʕal.ˈlam-t.#ni ʕal.ˈlam.-t#ak ʕal.ˈlam.-t#ək ʕal.ˈlam.-t#o ʕal.ˈlam-t.#(h)a ʕal.ˈlam-t.#na ʕal.ˈlam-t.#kom ʕal.ˈlam-t.#(h)om

Kenstowicz (2005) accounts for the unexpected realization of forms like /ʕallam-ət#o/ → [ʕal.la.ˈm-ə.t#o] ‘she taught him’ (as opposed to the expected *[ʕal.ˈlam.-t#o], which would have been homophonous with ‘I taught him / you-sg.m taught him’) by invoking a high-ranked constraint Paradigm Contrast, as illustrated in (15). Here the label Syncope is a stand-in for the constraint or constraints that effect syncope of schwa (for our purposes, a ban against [ə] in unstressed open syllables), and NonFinality requires that no foot be word-final, thus forcing antepenultimate stress in words ending in a sequence of three light syllables.

178

gunnar ólafur hansson

(15) Anti-homophony enforced: [ʕallaˈməto] ‘she taught him’ ̸= [ʕalˈlamto] ‘I taught him’ /ʕallam-ət#o/ a. b. c.

Paradigm Contrast

ʕal.(ˈla.mə).to ʕal.(ˈlam).to ʕal.la.(ˈmə.to)

Syncope

NonFinality

*!

*!

*

The strongest evidence in support of a homophony avoidance account of these facts is the observation that the same aberrant stress pattern is not seen in verbs which happen to have a different stem shape in their 3sg.f perfective forms than in the corresponding 1sg/2sg.m form. Thus, for example, so-called ‘hollow’ verbs like /ʃaːf/ ‘see’ display a vowel quality and length alternation within their subject-inflection paradigm: [ˈʃaː.f-et] ‘she saw’ but [ˈʃuf-t] ‘I saw/you-sg.m saw’. For this reason, there is no risk of homophony between object-inflected forms like ‘she saw him’ and ‘I saw him/yousg.m saw him’, and the former is free to surface with the regular stress pattern (and concomitant syncope), as shown in (16). (16) Anti-homophony irrelevant: [ˈʃaːf.-t#o] ‘she saw him’ ̸= [ˈʃuf.-t#o] ‘I saw him / you-sg.m saw him’ /ʃaːf-ət#o/ a. b. c.

(ˈʃaː.fə).to (ˈʃaːf).to ʃaː.(ˈfə.to)

Paradigm Contrast

Syncope

NonFinality

*! *

Kenstowicz (2005) adduces a number of similar examples from other modern Arabic vernaculars, where other deviations from the regular phonology are employed to maintain the same distinctions (3sg.f vs. 1sg/2sg.m before a vowel-initial object suffix). As the glosses above suggest, the 1sg and 2sg.m have become homophonous throughout the perfective paradigms in Syrian Arabic; both are formed with /-t/ (< Classical Arabic /-tu/ and /-ta/, respectively). This would appear to be a counterexample to the enforcement of Paradigm Contrast, in particular in light of the fact that the 1sg vs. 2sg.m distinction is maintained elsewhere in the language, such as in the imperfective (e.g. [ˈbək.tob] ‘I am writing’, [ˈbtək.tob] ‘you-sg.m are writing’). The 1sg vs. 2sg.m situation is not entirely parallel, however, in that these two paradigm cells are homophonous/syncretic already in their underlying representation. What is observed in (14), and other similar cases discussed by Kenstowicz (2005), is an apparent avoidance of derived instances of homophony/syncretism: ones that would otherwise have been expected to result from the regular application of phonological processes in the language. That being said, there is no obvious reason why high-ranked Paradigm Contrast could not equally well act as a trigger of (ad hoc) phonological

phonology

179

changes whose sole purpose would be to create a surface contrast between underlyingly homophonous paradigm cells. For example, one could easily imagine the 1sg vs. 2sg.m distinction being upheld on the surface by again resorting to stress shift, with /ʕallam-t/ → [ʕal.ˈlamt] (regular stress) in one case and /ʕallam-t/ → [ˈʕal.lamt] (aberrant stress) in the other. Whether this is a serious problem or not depends on exactly how Paradigm Contrast is defined formally. Kenstowicz (2005) does not provide an explicit definition, but refers to the one formulated by Rebrus and Törkenczy (2005), which is shown in (17).6 (17) Con [= Paradigm Contrast]

(Rebrus and Törkenczy 2005: 264)

The surface realizations of morphologically distinct members mi . . . mn of a paradigm x must be phonetically distinct. The key question here is what qualifies as being ‘morphologically distinct’. If the 1sg and 2sg.m differ at the level of morphosyntactic feature specifications, but involve exponents that are accidentally homophonous (both /-t/), that should presumably still count. Alternatively, one might imagine treating /-t/ as being underspecified for the first vs. second person opposition ([±addressee]), as well as for gender (whereas 2sg.f /-ti/ would be contrastively specified as [+addressee, +feminine], for example). In that case it could be argued that there is no 1sg vs. 2sg.m ‘morphological distinction’ for Paradigm Contrast to uphold in the Syrian Arabic perfective—rather, the two constitute a single syncretic paradigm cell, represented by a single output form. Such problems of formal implementation aside, the very validity of intraparadigmatic homophony avoidance as a synchronic guiding principle in morphophonology is a matter of some doubt. As in the case of paradigm uniformity (see Section 8.3.1 above), a number of researchers have expressed scepticism in this regard; for a recent critique, see Mondon (2009). Several recent works attempt instead to define a principled mechanism whereby anti-homophony effects can arise through blocking (or apparent exceptions to) otherwise regular historical sound changes (Blevins 2005; Blevins and Wedel 2009; Gessner and Hansson 2005; Mondon 2009). However, this does still leave open the possibility that there is some role to play for homophony avoidance in online morphological processing and production, which could then provide a potential source for patterns that can in turn become grammaticalized. For the potential role of intra-paradigmatic homophony avoidance as a contributing factor in the emergence of paradigm gaps, see Baerman (2011, this volume) and Section 8.3.3. 6 Rebrus and Törkenczy (2005: 266) further propose that constraints on paradigm contrast and uniformity be relativized to particular feature dimensions (‘Con(D): A form realizing some value of a morphosyntactic dimension D of paradigm x must be phonetically distinct from forms realizing other values of D’). This seems irrelevant in the Syrian Arabic case, as the same dimensions (person, gender) are involved in both the avoided 3sg.f vs. 1sg/2sg.m identity and the tolerated 1sg vs. 2sg.m identity.

180

gunnar ólafur hansson

8.3.3 Phonotactics, Alternations, and Paradigm Gaps A final area where phonology intersects with paradigm structure is in the phenomenon of paradigm gaps (Hetzron 1975; Iverson 1981; Hansson 1999; Albright 2003; Raffelsiefen 2004; Pertsova 2005; Sims 2006, 2009; Rice and Blaho 2009; Baerman et al. 2010; Baerman 2011, this volume). In many cases of defective paradigms, the occurrence and distribution of gaps within the paradigm is strongly correlated with certain aspects of the phonological shape of the expected (but non-existent) word forms in question. Phonologically motivated paradigm gaps can be said to come in two flavours: ones that seem primarily driven by phonotactic well-formedness considerations, and ones that are lexically arbitrary but appear motivated by morpho-phonological alternation patterns. (However, this dichotomy may be illusory; see Albright 2009: 158–60 for discussion.) A striking example of what appears to be a phonotactically motivated paradigm gap occurs for imperatives in Norwegian, as described by Rice (2007). Imperatives of verbs consist of the bare root, as shown in (18a). However, for verb stems that end in a rising-sonority consonant cluster—one which cannot be parsed as a complex coda—no imperative form exists (at least for certain dialects, Rice 2003), as indicated in (18b). This is a slight oversimplification, however, in that it is only in utterance-final or pre-consonantal contexts that the indicated forms are unacceptable. Should the imperative be embedded in a clausal context in which the next word happens to be vowel-initial, the imperative will surface intact, its final cluster resyllabified as a coda-onset sequence; this is illustrated in (19)–(20) with examples from Rice (2007). (18) Norwegian: imperative gaps due to syllable structure infinitive imperative a.

spiːs-ə snakː-ə løft-ə

‘eat’ ‘talk’ ‘lift’

spiːs snakː løft

‘eat!’ ‘talk!’ ‘lift’

b.

klatr-ə sykl-ə padl-ə oːpn-ə

‘climb’ ‘bike’ ‘paddle’ ‘open’

∅ (*klatr) ∅ (*sykl) ∅ (*padl) ∅ (*oːpn)

‘climb!’ ‘bike!’ ‘paddle!’ ‘open!’

(19) a. * Sykl ned bakken! bike.imp down hill.def ‘Bike down the hill!’ b.

Sykl opp bakken! bike.imp up hill.def ‘Bike up the hill!’

(Rice 2003, 2007)

phonology

181

(20) a. *Ikke klatr p˚a møblene! not climb.imp on furniture.pl.def ‘Don’t climb on the furniture!’ b.

Klatr ikke p˚a møblene! climb.imp not on furniture.pl.def ‘Don’t climb on the furniture!’

c.

opp p˚a møblene! Ikke klatr not climb.imp up on furniture.pl.def ‘Don’t climb up onto the furniture!’

Their dependence on the phrasal phonological context clearly suggests that the paradigm gaps in Norwegian imperatives are a relatively ‘late’ phenomenon, derivationally speaking. Another notable fact about these gaps is that the very same kinds of sonority sequencing violations are subject to repair, by schwa epenthesis, in the nominal morphology. Thus, for example, the unsuffixed singular form of the noun /sykl/ ‘bicycle’ (pl [sykl-ər]) is realized as disyllabic [sykːəl], with epenthesis to avoid the word-final /kl/ cluster.7 For this reason, uncertainty about whether or not to apply epenthesis in the imperatives in (18b) might be a contributing factor as well. The second type of paradigm gap in which phonological factors also seem to be implicated targets inflected forms which, had they existed, might have been expected to exhibit some phonological alternation (typically in the stem) relative to other forms in the paradigm (Hetzron 1975; Albright 2003, 2009). The well-known cases of gaps in verb paradigms of Spanish and Russian are a prime example of this. In Spanish, gaps target exactly those forms in the present indicative paradigm in which raising or diphthongization of the stressed stem vowel, or appearance of a velar stop (typically after an otherwise stem-final [θ]/[s], [l] or [n]), is observed in a large number of verbs of the corresponding inflectional classes (21). In Russian, gaps in the nonpast paradigm similarly target exactly the form in which palatalization (postalveolarization, rather) of the stem-final coronal obstruent would be expected to occur (22). (21) Spanish: gaps in present indicative paradigms

1sg 2sg 3sg 1pl 2pl 3pl 7

(Albright 2003, 2009)

/abol-ir/ ‘abolish’

/sol-er/ ‘be used to’

/as-ir/ ‘grasp’

/kreθ-er/ ‘grow’

— — — aβoˈl-imos aβoˈl-is —

ˈswel-o ˈswel-es ˈswel-e soˈl-emos soˈl-eis ˈswel-en

— ˈas-es ˈas-e aˈs-imos aˈs-is ˈas-en

ˈkreθk-o ˈkreθ-es ˈkreθ-e kreˈθ-emos kreˈθ-eis ˈkreθ-en

For an explanation of the medial gemination in [sykːəl] ‘bicycle’, see Rice (2006).

182

gunnar ólafur hansson

(22) Russian: gaps in nonpast paradigms

1sg 2sg 3sg 1pl 2pl 3pl

(Baerman 2008; Albright 2009)

/u-bed-itj / ‘convince’

/sled-itj / ‘follow’

/o-bez-les-itj / ‘deforest’

/gas-itj / ‘extinguish, quench’

— ubeˈdj -iʃ ubeˈdj -it ubeˈdj -im ubeˈdj -itj e ubeˈdj -at

slj eˈʒ-u sleˈdj -iʃ sleˈdj -it sleˈdj -im sleˈdj -itj e sleˈdj -at

— obezˈlesj -iʃ obezˈlesj -it obezˈlesj -im obezˈlesj -itj e obezˈlesj -at

gaˈʃ-u ˈgasj -iʃ ˈgasj -it ˈgasj -im ˈgasj -itj e ˈgasj -at

Several of the examples of defective paradigms adduced by Baerman (this volume) fit this description, such as where the Finnish verb stem /erka(n)-/ ‘spread out’ ([erkan-eː] ‘s/he spreads out’) is missing all forms that would normally have called for a weakgrade stem alternant [era(n)-], including the infinitive *[era-ta] (cf. [vaiken-eː] ‘s/he is quiet’, [vaie-ta] ‘to be quiet’). Similarly, as Baerman points out, the gen.pl gap in the paradigm of Russian /meʧt-a/ ‘dream’ (expected gen.pl form: *[ˈmeʧt]) is at least partly due to the fact that this would be the only form in the paradigm to exhibit stem stress, as all other forms carry suffix stress: nom.sg [meʧˈt-a], gen.sg and nom.pl [meʧˈt-ɨ], and so forth (cf. the phonotactically equivalent and non-gapped /maʧt-a/ ‘mast’, for which the gen.pl [ˈmaʧt] receives support from stem-stressed forms like nom.pl [ˈmaʧt-ɨ]). The most convincing explanation for paradigm gaps of this type is that of Albright (2003, 2009), in which uncertainty on the part of the speaker is the driving factor: ‘gaps occur when speakers know that an inflected form must stand in a certain relation to another inflected form, but the language does not provide enough data to be certain of what that relation should be’ (Albright 2009: 118). For example, the Spanish stemvowel alternations seen in (21) are not observed in all verbs belonging to the relevant inflectional classes (compare [soˈl-er] ‘be used to’, 1sg [ˈswel-o], to [koˈm-er] ‘eat’, 1sg [ˈkom-o]). The lexical distribution of defective verbs, and speakers’ confidence in their own production of putatively gapped inflected forms, appear to be intimately correlated with the statistics of the lexicon—namely, how prevalent alternation happens to be for particular inflectional (sub)classes, or for verb stems with particular segmental properties. Since they typically involve alternations in the phonological shape of the stem, gaps of this kind have the effect of producing paradigms that are ‘uniform’ (across their non-gapped forms) in the sense of Section 8.3.1. Avoidance of paradigm-internal homophony or syncretism (Section 8.3.2) also appears to play a contributing role in some cases (Baerman 2011). For example, in the large and productive inflectional class of weak feminine nouns in Icelandic (23), many nouns have a gap in the gen.pl. The historically inherited gen.pl ending for nouns of this class is /-na/.8 However, some 8 Aside from weak feminine nouns, the only other inflectional class that retains /-na/ are the weak neuters, a small set of half a dozen nouns mostly denoting body parts (e.g. /øyɣ-a/ ‘eye’, /nir-a/ ‘kidney’).

phonology

183

weak feminines have adopted the /-a/ gen.pl ending that is used in virtually all other inflectional classes, whereas others vacillate between the two or have an outright gap (with considerable inter- and intra-speaker variation as to whether a given noun can be used in the gen.pl, and if so, with which ending). (23) Icelandic: gaps in weak feminine noun paradigms

(cf. Baerman 2011)9

/kh ul-a/ ‘bullet’

/ch ɪst-a/ ‘coffin’

/ph ɛr-a/ ‘pear’

/hɔl-a/ ‘hole’

/fɪðl-a/ ‘violin’

sg nom acc dat gen

kh uːl-a kh uːl-ʏ kh uːl-ʏ kh uːl-ʏ

ch ɪst-a ch ɪst-ʏ ch ɪst-ʏ ch ɪst-ʏ

ph ɛːr-a ph ɛːr-ʏ ph ɛːr-ʏ ph ɛːr-ʏ

hɔːl-a hɔːl-ʏ hɔːl-ʏ hɔːl-ʏ

fɪðl-a fɪðl-ʏ fɪðl-ʏ fɪðl-ʏ

pl nom acc dat gen

kh uːl-ʏr kh uːl-ʏr kh uːl-ʏm kh ul-na

ch ɪst-ʏr ch ɪst-ʏr ch ɪst-ʏm ch ɪst-na

ph ɛːr-ʏr ph ɛːr-ʏr ph ɛːr-ʏm ph ɛːr-a

hɔːl-ʏr hɔːl-ʏr hɔːl-ʏm —

fɪðl-ʏr fɪðl-ʏr fɪðl-ʏm —

As can be seen from the paradigms in (23), /-na/ is the only inflectional suffix that is consonant-initial. As such, it can trigger a number of regular phonological processes that have the effect of creating an alternation in the surface shape of the stem in the gen.pl form vis-à-vis all others in the paradigm. These include vowel shortening ([suːl-a] ‘pillar’, gen.pl [sul-na]), hardening ([kh raːv-a] ‘demand’, gen.pl [kh rap-na]), depalatalization ([ɛhc-a] ‘widow’, gen.pl [ɛhk-na]), preaspiration ([vɪːkh -a] ‘week’, gen.pl [vɪhk-na]), cluster simplification ([th uŋk-a] ‘tongue’, gen.pl [th uŋ-na]), and stop excrescence ([viːs-a] ‘quatrain’, gen.pl [vist-na]). Avoidance of these sorts of alternations is no doubt a primary motivation for avoiding /-na/ in favour of /-a/. For example, stems ending in a singleton palatal stop, for which /-na/ would trigger a combination of depalatalization, vowel shortening, and preaspiration, tend very strongly toward either /-a/ or a gap (e.g. [θɛːch -a] ‘roof, cover’, gen.pl *[θɛhk-na], ?[θɛːch -a]). By contrast, stems ending in a geminate and/or preaspirated palatal, or a cluster involving a palatal, are typically much more tolerant of /-na/, since here depalatalization is the only consequence ([ch ɪr̥c-a] ‘church’, gen.pl [ch ɪr̥k-na]; cf. also [ɛhc-a] ‘widow’, just mentioned). In other cases the avoidance of /-na/ appears to be phonotactically motivated to some extent, in that it would lead to a consonant cluster that is otherwise unattested (and for which a choice of phonological repair is not immediately obvious), such as [th ʏðr-a] ‘pouch’, gen.pl *[th ʏðr-na]. 9

Baerman (2011) uses hola ‘hole’ as his illustrative example of a defective weak feminine noun. A Google search produces a considerable number of hits for each of hol-na and hol-a in gen.pl contexts, suggesting that for at least some speakers this word does not have a gapped paradigm. For the other gapped noun shown in (23), fiðla ‘violin’, both gen.pl fiðl-na and fiðl-a are markedly rare. (In the definite form, gen.pl fiðl-a-nna is robustly attested, lending furter support to a homophony avoidance account; the addition of the suffixed definite article disambiguates the gen.pl from nom.sg fiðl-a-n.)

184

gunnar ólafur hansson

Whereas the dispreference for the /-na/ allomorph is thus arguably phonologically motivated, at least in part, what underlies the dispreference for /-a/ seems to be the fact that it renders the gen.pl homophonous with the (much more frequent and salient) nom.sg form (Baerman 2011). (Indeed, as Baerman points out, it is hardly an accident that the two inflectional classes of nouns that have retained gen.pl /-na/ in the first place are precisely the ones in which the nom.sg happens to be marked with /-a/.) The conflict resulting from these two opposing desiderata, then, is what drives speakers’ uncertainty about the gen.pl form of many weak feminines, a paradigm gap being the most extreme manifestation of such uncertainty. Frequency is presumably a third contributing factor exacerbating the uncertainty: the relatively low token frequency of gen.pl forms in general, and the low type frequency of /-na/ as an exponent of gen.pl across the lexicon as a whole.

8.4 Phonologically Conditioned Allomorphy ............................................................................................................................................................................. The most obvious way in which the phonology of a language can make its presence felt in the inflectional system is by giving rise to alternations in the sound shape of some morpheme, stem, or lexeme. For example, Modern Icelandic lengthens vowels in stressed open syllables, and hardens certain voiced fricatives to (voiceless unaspirated) stops before [l, n], and this results in a length and manner alternation in the stem of [saːɣ-a] ‘story-nom.sg’ vs. [sak-na] ‘story-gen.pl’. Similarly, the regular English plural marker surfaces as [-z], [-s], or [-ɨz] depending on whether the preceding segment is voiced or voiceless, sibilant or non-sibilant. Whenever such an alternation can be related directly to some generalization about phonotactics, allophonic distribution, neutralization patterns, or the like—where that generalization is valid either across the board or in some well-defined set of morphological contexts—the standard assumption in generative phonology (and to some extent pre-generative frameworks as well) is that the alternation is to be factored out by positing a single, underlying (input, phonemic, lexical) representation of the morpheme in question. Thus the regular English plural is simply /-z/ as far as the morphological system is concerned, regardless of the fact that this /-z/ happens sometimes to surface as [-s] or [-ɨz] due to the exigencies of English segmental phonology. Instances of alternation do however exist which, although phonologically conditioned, do not seem amenable to this kind of analysis. There can be several possible reasons (not mutually exclusive) for why this may be the case. First, the alternants may be too divergent from each other to be derivable from a single phonological representation. Second, it may be impossible to derive the differences in sound shape from the different conditioning environments in which the alternants occur. Third, the alternation may be attested too sporadically to warrant treatment as the product

phonology

185

of a (perhaps construction-specific) phonological process. Alternations that fail on one or more of these criteria make up the important class of morphology–phonology interface phenomena that are known collectively as phonologically conditioned allomorphy (Carstairs 1988; Nevins 2011). Since the term ‘allomorphy’ is occasionally used in a wider sense, encompassing purely phonological alternations such as English [-z]∼[-s]∼[-ɨz] as well, the less ambiguous term phonologically conditioned suppletive allomorphy is preferred by some. A textbook example of phonologically conditioned allomorphy is the English indefinite article, which is [ə] before a consonant but [ən] before vowels. No other English morpheme displays that kind of [n]∼Ø alternation, and there is no independently attested process of either nasal deletion or nasal epenthesis that the alternation could be attributed to. Hence we are forced to admit that the indefinite article has two phonological representations, two allomorphs, /ə/ and /ən/, whose relative distribution is governed by phonological factors. A more subtle case is the voicing alternation exhibited by a handful of suffixes in Basque, discussed by Mascaró (2007). Modern Basque does not have a productive process of postnasal voicing ([mendi] ‘mountain’, [kontu] ‘count’, [hon-tas̻] ‘this-instr’). However, a small number of inflectional and derivational suffixes display an alternation [t]∼[d] or [k]∼[ɡ], with the voiced alternant systematically appearing after nasal-final bases (and to some extent after /l/-final ones as well, subject to dialect variation). This is illustrated in (24) with the example of ablative [-tik]∼[-dik]. (24) Basque: postnasal voicing allomorphy bilbo ‘Bilbao’ iɾun ‘Irun’ (place name) non ‘where’

(Mascaró 2007)

bilbo-tik ‘from Bilbao’ iɾun-dik ‘from Irun’ non-dik ‘from where’

As a phonological process, post-nasal voicing is both phonetically natural and typologically frequent, and is commonly analysed as resulting from a high-ranked, and phonetically grounded, markedness constraint *NC̥ (Hayes 1999; Pater 2001). If those affixes of Basque that show the voicing alternation in (24) formed a coherent class on independent grounds (morphological, syntactic, or semantic), then that would invite the possibility of interpreting this as a case of construction-specific phonology. For example, the affixes in question could be seen as belonging to a cophonology or stratum in which *NC̥ is ranked higher than otherwise; alternatively, they could be indexed to a specific and low-ranked version of the faithfulness constraint that demands preservation of underlying [±voice] values. Mascaró (2007) argues that this is not tenable in the Basque case. The suffixes displaying the voicing alternation constitute a heterogeneous set of about a half-dozen morphemes, some of which are clearly derivational (such as [-tar]∼[-dar], e.g. [bilbo-tar] ‘Bilbaoese’, [iɾun-dar] ‘Irunese’), some inflectional (such as the ablative shown in example (24) and the future participle suffix [-ko]∼[-ɡo]), and one is even a clause-final clitic ([=ta]∼[=da], an enclitic version of the conjunction /eta/ ‘and; since’).

186

gunnar ólafur hansson

It would appear, then, that the observed voicing alternations in (24) cannot be construed as the product of a phonological process in the usual sense. Instead, the existence of two alternative surface shapes [-tik] and [-dik] must to some extent be relegated to the morphology and/or the lexicon. From the point of view of the phonology, then, the Basque ablative suffix has two allomorphs, /-tik/ and /-dik/, which despite their similarity are, for all intents and purposes, suppletive. From an item-based perspective, the (lexical, underlying) phonological representation of the ablative suffix morpheme consists of the set /{-tik, -dik}/. From a realizational perspective one might posit two separate suffixation rules, involving the mappings X → Xtik and X → Xdik, respectively. The difference is not important here, but for simplicity and consistency with most of the relevant phonological literature the discussion in the following subsections is couched in item-based (morpheme-based) terms.

8.4.1 Allomorph Selection as Optimization Even though the Basque alternation just described is not itself due to a phonological process, the distribution of the allomorphs is nevertheless conditioned by phonological factors. In Optimality Theory, the standard approach to such phonologically conditioned allomorphy has been to model allomorph selection as output optimization (Mester 1994; Kager 1996; Mascaró 1996; Rubach and Booij 2001). Because both phonological shapes are already available in the input, the phonological grammar is free to choose that output shape which is less marked (phonologically) in the environment in question, without the cost of violating input–output correspondence. In this way, low-ranked markedness constraints that are otherwise not enforced in the language are able to make their presence felt: an ‘emergence of the unmarked’ effect (McCarthy and Prince 1994). The Basque state of affairs can be seen as reflecting a ranking Ident[±voice]-IO ≫ *NC̥ ≫ *VoicedObstruent, where the faithfulness constraint Ident[±voice]-IO penalizes any deviations from input voicing specifications, and *VoicedObstruent is a context-free markedness constraint encoding the generally marked status of voiced vis-à-vis voiceless obstruents. In cases without allomorphy, such as root-internal /nt/ in /kontu/ ‘count’ or the non-alternating instrumental /-tas̻/ in /hon-tas̻/ ‘thisinstr’, the undominated constraint Ident[±voice]-IO will guarantee that the stop remains voiceless into the output, despite violating lower-ranked *NC̥. When the input representation of a morpheme contains more than one allomorph, however, as in /iɾun-{tik, dik}/ ‘Irun-abl’, it is possible to satisfy *NC̥ for free, as it were, without violating Ident[±voice]-IO. This is illustrated by the tableaux in (25)–(27). Following Mascaró (2007), subscript indices signal which of the input allomorphs is being referenced (for purposes of faithfulness evaluation) in the output candidate in question. Thus [iɾun-dik1 ] is an unfaithful rendering of /iɾun-tik1 / (with post-nasal voicing enforced, analogous to the losing candidate in (25b) *[hon-das̻]), whereas the homophonous [iɾun-dik2 ] is a fully faithful rendering of /iɾun-dik2 /.

phonology

187

(25) /hon-tas̻/ → [hon-tas̻] ‘this-instr’ /hon-tas̻/ a. b.

Ident[±voi]-IO

hon-tas̻ hon-das̻

*NC̥

VoiObs

*

*!

*

(26) /iɾun-{tik, dik}/ → [iɾun-dik] ‘Irun-abl’ /iɾun-{tik1 , dik2 }/ a. b. c. d.

iɾun-tik1 iɾun-tik2 iɾun-dik1 iɾun-dik2

Ident[±voi]-IO *! *!

*NC̥ *! *

VoiObs

* *

(27) /bilbo-{tik, dik}/ → [bilbo-tik] ‘Bilbao-abl’ /bilbo-{tik1 , dik2 }/ a. b. c. d.

bilbo-tik1 bilbo-tik2 bilbo-dik1 bilbo-dik2

Ident[±voi]-IO *! *!

*NC̥

VoiObs

* *!

In a case like the Basque one, there is an obvious and natural connection between the conditioning environment (post-nasal vs. elsewhere) and the substantive difference between the two alternants (voiced vs. voiceless stop morpheme-initially). This invites an alternative solution. Rather than represent the suffix as a set, or disjunction, of two phonological shapes /{-tik, -dik}/, we could instead posit a single phonological representation that unifies the two by factoring out their featural differences: /-Tik/, where /T/ stands for a coronal stop that is unspecified for [±voice]. Inkelas (1995) proposes exactly this sort of archiphonemic underspecification analysis of the lexical contrast between alternating and non-alternating stem-final obstruents in Turkish. As long as the voiced–voiceless opposition is seen as equipollent rather than privative, and as long as (some) faithfulness constraints penalize feature-value changes over and above feature-value insertion, this will allow for an analysis along similar lines as above, with Faith-IO ≫ *NC̥ ≫ *VoiObs. Exactly as with the set/disjunction approach, the underspecification approach encodes the difference between alternating (allomorphic) and non-alternating (non-allomorphic) morphemes in their lexical representation, but the two deal with the alternating ones in opposite ways. The former encodes both feature values at once in the input representation ([−voice] in /-tik/, [+voice] in /-dik/), while the latter encodes neither ([0voice] in /-Tik/). While an archiphonemic underspecification approach could arguably handle the voicing allomorphy in Basque, it runs aground on any of the numerous cases of phonologically conditioned allomorphy where the alternating morpheme shapes cannot be

188

gunnar ólafur hansson

construed as anything but suppletive. A few illustrative examples are shown in (28). The first three cases involve a post-consonantal vs. post-vocalic alternation, which is often best understood in terms of syllable structure. Thus, for example, the Korean forms in (28b) both have CV.CV shape, whereas the exact opposite distribution of the two allomorphs would have resulted in the more marked structures CVC.CV and CV.V. Note that in the Udihe case in (28d), the laryngealization of the stem-final vowel (e.g. [zawa˷] from /zawa/ ‘take, grab’) can be interpreted as a floating [+constricted glottis] featural suffix. (28) a. Moroccan Arabic 3rd singular masculine (Harrell 1962; Mascaró 1996) [-u] [-h]

/ C ___ / V ___

[kta.b-u] [xtʕ a-h]

b. Korean nominative [-i] [-ka]

/ C ___ / V ___

c. Djabugay genitive singular 2001) [-ŋun] [-n]

/ C ___ / V ___

d. Udihe perfective [-ɡe] [-V˷ ]

/ V[+high] ___ / V[−high] ___

‘his book’ ‘his error’ (Carstairs 1988; Embick 2010)

[pa.p-i] [kh o.-ka]

‘cooked.rice-nom’ ‘nose-nom’

(Patz 1991; Kager 1996; Rubach and Booij [ɡa.ɲal.-ŋun] [ɡu.lu.du-n]

‘goanna-gen’ ‘dove-gen’

(Nikolaeva and Tolskaya 2001; Bye 2007) [doɡdi-ɡe] [zawa˷]

‘hear-pfv’ ‘take-pfv’

In some cases, synchronically suppletive allomorphy has its historical origins in nonsuppletive phonological alternations. For example, both Moroccan Arabic [-h] and [-u] in (28a) are reflexes of Classical Arabic /-hu/ by historical processes of apocope and lenition/deletion.

8.4.2 The Problem of Default Allomorphs In the Basque case in examples (24)–(27), the distribution of each of the two allomorphs can be attributed to well-known phonological factors. The context-free markedness constraint *VoicedObstruent favours [-tik] over [-dik] and so has the effect of defining the former as the default or elsewhere allomorph, whereas in the environment N___ the higher-ranked context-sensitive markedness constraint *NC̥ becomes relevant and favours the latter, special-case allomorph. Similarly in the Moroccan Arabic case in (28a), uncontroversial syllable-structure constraints will penalize both the use of [-u] after vowel-final bases (creating hiatus, an Onset violation, *[xtʕ a.-u]) and the use of [-h] after consonant-final bases (creating a

phonology

189

complex coda, a *Complex violation, *[ktab-h]). The complementarity of the two allomorphs is thus correctly captured. However, when one allomorph appears to serve as an elsewhere case, it is often unclear how its default status is to be derived, and in many cases outright stipulation seems required. In the Djabugay genitive in (28c), it is understandable why [-ŋun] would be favoured over [-n] after consonant-final bases ([ɡa.ɲal.-ŋun] ‘goanna-gen’ avoids the complex coda of its competitor *[ɡa.ɲal-n]), but it is less obvious exactly what favours [-n] over [-ŋun] otherwise (e.g. [ɡu.lu.du-n] ‘dove-gen’, not *[ɡu.lu.du.ŋun]). Kager (1996) encodes that preference by stipulating an obviously parochial and language-specific constraint Genitive=/-n/, but Rubach and Booij (2001) attribute it to the relatively marked status of dorsal place over coronal place, and hence of the velar nasal [ŋ] of the [-ŋun] alternant. In the Udihe case in (28d), the avoidance of laryngealization on a stem-final high vowel has an obvious phonological motivation in the fact that laryngealized high vowels are categorically absent from the language, but it is far from clear why suffixation of [-ɡe] should be disfavoured after other vowels. If the perfective suffix has the allomorphic representation /{-ɡe, [+constricted glottis]}/, why is *[zawa-ɡe] worse than [zawa˷] as the perfective of /zawa/ ‘take, grab’? To address this type of problem, Mascaró (2007) borrows from morphological theory the notion of morphological defaults (cf. Brown and Hippisley 2012; Brown, this volume), and incorporates it into the way listed allomorphs are represented and manipulated in the phonology. Mascaró proposes that (lexical, underlying) phonological representations involving allomorphy be construed as partially ordered sets, where the ordering of allomorphs encodes their relative priority. The ‘first’ allomorph is the default or elsewhere case, others are special cases limited to particular contexts. Selection of a lower-priority allomorph entails violating a special faithfulness constraint called Priority, and can thus only be optimal if it serves as a means of satisfying some higher-ranked constraint. Couched in Mascaró’s framework, the alternation between laryngealization of a stem-final vowel and suffixation of [-ɡe] in Udihe perfectives might be captured by representing the perfective suffix as the ordered set /{[+constricted glottis] > -ɡe}/. High-ranked Priority ensures that, other things being equal, laryngealization is the preferred option (despite the unfaithfulness that this causes in the stem-final vowel). However, the even higher-ranked ban against laryngealized high vowels, *[+high, +constricted glottis], forces the emergence of the lower-priority allomorph [-ɡe] if and only if the stem-final vowel is high. This is shown in (29)–(30). (29) /zawa-{[+constricted glottis] > -ɡe}/ → [zawa˷] ‘take-pfv’ /zawa-{[+c.g.]1 > -ɡe2 }/ a. b.

zawa˷1 zawa-ɡe2

*[+high, +c.g.]

Priority *!

Ident[±c.g.]-IO *

190

gunnar ólafur hansson

(30) /doɡdi-{[+constricted glottis] > -ɡe}/ → [doɡdi-ɡe] ‘hear-pfv’ /doɡdi-{[+c.g.]1 > -ɡe2 }/ a. b.

*[+high, +c.g.]

doɡdi˷1 doɡdi-ɡe2

*!

Priority *

Ident[±c.g.]-IO *

8.4.3 Non-optimizing Allomorphy Extrinsic ordering among allomorphs addresses cases in which the selection of one of the allomorphs cannot easily be construed as phonologically optimizing. However, there exist numerous examples of phonologically conditioned allomorphy where neither/none of the allomorphs appears to involve optimization. One such case is Kaititj (Koch 1980; Paster 2005), where the ergative/instrumental/locative suffix is realized as [-ŋ] after disyllabic stems but as [-l] after stems of more than two syllables, as illustrated in (31).10 (31) Kaititj: syllable-counting /-ŋ/ ∼ /-l/ allomorphy aˈki-ŋ ‘head-erg/instr/loc’ ilˈtj i-ŋ ‘hand-erg/instr/loc’ ajnˈpi-ŋ ‘pouch-erg/instr/loc’

(Paster 2005)

aˈliki-l ‘dog-erg/instr/loc’ aˈʈuji-l ‘man-erg/instr/loc’ aˈɣirki-l ‘sun-erg/instr/loc’

Other cases abound in which the connection between the shape of the allomorphs and the phonological factors controlling their distribution seems equally arbitrary, or even anti-optimizing. In Korean (Lapointe 2001), the conjunctive suffix is realized as [-wa] after vowels but [-kwa] after consonants, even though the latter results in the combination of a coda and complex onset (e.g. [mom.-kwa] ‘body-conj’) that could easily have been avoided by choosing the former allomorph instead (*[mom.-wa] or *[mo.m-wa]). A particularly striking case is that of the postposed definite article in Haitian Creole (Nikiema 1999; Klein 2003; Paster 2006; Bonet et al. 2007; Bye 2007; Embick 2010), which is either [-la] or [-a] depending on the phonological properties of the preceding noun. If the noun ends in a consonant, [-la] is used (32a). If the noun ends in a vowel, [-a] is used (32b–c), and if the final vowel is [+Advanced Tongue Root] a homorganic glide is inserted to break the hiatus between the two vowels (32c).11 10 In Kaititj /l/ → [ɭ] when the preceding consonant is apical. The allomorph /-l/ may thus surface as [-ɭ] in some cases (e.g. [aˈɣiri-ɭ] ‘kangaroo-erg’). 11 There appears to be some variation as to whether the glide insertion pattern is confined to [+ATR] vowels or holds after [ε ɔ] as well (see Klein 2003 for discussion); the pattern depicted in (32) is the one described by Valdman (1978). Also, phonological processes of nasalization and assimilation produce other surface alternants not shown here, e.g. [mãɡ-lã] ‘mango-def’, [madam-nã] ‘lady-def’, [tiɡasõ[w]-ã] ‘boy-def’ (Nikiema 1999).

phonology (32) Haitian Creole: anti-optimizing /-la/ ∼ /-a/ allomorphy a.

liv ma.lad ba.ɡaj

‘book’ ‘sick’ ‘thing’

liv.-la ma.lad.-la ba.ɡaj.-la

‘book-def’ ‘sick.person-def’ ‘thing-def’

b.

pa.pa bɔ.kɔ vε

‘father’ ‘sorcerer’ ‘glass’

pa.pa.-a bɔ.kɔ.-a vε.-a

‘father-def’ ‘sorcerer-def’ ‘glass-def’

c.

la.pli pa.pje tu

‘rain’ ‘paper’ ‘hole’

la.pli.[j]-a pa.pje.[j]-a tu.[w]-a

‘rain-def’ ‘paper-def’ ‘hole-def’

191

(Bye 2007)

The hiatus in (32b) could easily have been avoided by selecting the [-la] allomorph instead (*[pa.pa.-la]). As Bye (2007) points out, the choice of allomorph is particularly surprising in (32c), given that hiatus is in fact being actively avoided there but by means of a phonological operation. Using the other allomorph would have been a less costly solution than epenthesis in terms of faithfulness (e.g. *[pa.pje.-la]). Finally, in (32a) the [-a] allomorph would have resulted in a less marked syllable structure by avoiding the noun-final coda (*[li.v-a]), although at the expense of misalignment between syllable and morpheme boundaries (Klein 2003; Bonet et al. 2007; see Section 8.4.4).

8.4.4 Subcategorization vs. Optimization The pervasive occurrence of (apparently) non-optimizing cases of phonologically conditioned allomorphy like the Haitian Creole one has prompted some researchers (Paster 2005, 2006, 2009; Bye 2007; Embick 2010) to reject altogether the notion of allomorph selection as output optimization. On the basis of an extensive typological survey of phonologically conditioned allomorphy, Paster (2006, 2009) argues that the observed typology conforms much better to the predictions of an analysis in terms of subcategorization than to those of the optimization approach. In a subcategorization approach, the Kaititj /-ŋ/ ∼ /-l/ allomorphy in (31) could be expressed in terms of the two subcategorization frames shown in (33), adopting the notation of Paster (2006). Since the frames in (33) stand in a subset relation, the Elsewhere Condition or P¯an.ini’s principle (Kiparsky 1973, etc.) dictates that where relevant, (33a) will take precedence over (33b), exactly as desired. (33) Subcategorization frames for ergative (etc.) allomorphy in Kaititj a.

[ [ # σ σ # ]stem /-ŋ/ erg sfx ]erg word

b.

[[

]stem /-l/ erg sfx ]erg word

192

gunnar ólafur hansson

A fundamental difference between the subcategorization approach and the output optimization approach is that the former assumes that morphology (here: allomorph selection) ‘precedes’, and hence feeds, phonology, whereas in the latter phonology and morphology operate in parallel. Under output optimization, information flows in both directions (M↔P). It is thus possible for the choice of allomorph to be influenced by the eventual outcome of whatever phonological processes might be triggered by the allomorph selection itself (an instance of what Bakovi´c 2007 dubs ‘counterfactual derivation’). Indeed, this is more than a possibility; it is what is predicted to occur, other things being equal. One of the arguments that Paster (2006, 2009) adduces for subcategorization over output optimization is that this sort of crucially outputbased allomorph selection appears to be unattested, whereas clear cases of crucially input-based selection do exist. For example, in Turkish the third person singular possessive suffix has two allomorphs (each of which is subject to vowel harmony), [-i]/[-y]/[-ɯ]/[-u] after consonants and [-si]/[-sy]/[-sɯ]/[-su] after vowels. However, Turkish also has a systematic process of Velar Deletion, whereby /k/ → Ø / V___V (Sezer 1981; Inkelas and Orgun 1995), and this creates a potential dilemma for allomorph selection in the case of /k/final nouns. For a noun like /bebek/ ‘child’, subcategorization predicts that the input representation that is supplied to the phonology will be /bebek-I/ ‘child-3sg’ which, due to Velar Deletion, will in turn surface as [be.be.-i]. This is indeed what we find. However, with the /k/ gone, from an output perspective the base is now vowel-final, not consonant-final, and this should have made the allomorph /-I/ equally inappropriate as it is after genuinely vowel-final nouns. Output optimization would thus predict selection of /-sI/ as a means of avoiding this situation, as illustrated in (34). Max-IO is the Faithfulness constraint penalizing deletion; we know that this constraint outranks NoCoda in Turkish because closed syllables are not generally repaired by means of consonant deletion. The sad-face symbol indicates the candidate representing the actual Turkish output form, while the pointing hand designates the incorrectly predicted output, *[be.bek.-si]. (34) /bebek-{I, sI}/ → [be.be.-i] ‘child-3sg’ /bebek-{I1 , sI2 }/ a. b. c. d.

be.be.k-i1 be.be.-i1 be.bek.-si2 be.be.-si2

*VkV *!

Max-IO *! *!

NoCoda

*

Onset *

In other words, allomorph selection in Turkish 3sg /{-I, -sI}/ is rendered opaque by the phonological process of Velar Deletion, and this is exactly as predicted by a model in which allomorph selection logically precedes (i.e. feeds) the phonological input– output mapping rather than taking place as part of that mapping itself. Paster’s (2006) point is not that the opacity as such is a problem for the optimization approach,

phonology

193

but rather the typological generalization that allomorph selection is always opaque (input-based) in such cases, never transparent (output-based). Another problem with the output optimization approach is that it is rather unconstrained in practice; whether a given pattern of allomorph distribution can be construed as phonologically optimizing or not seems limited only by the ingenuity of the analyst. Considerable freedom is created if one of the allomorphs can be freely designated as default (see Section 8.4.2). Another source of descriptive power are alignment constraints that can regulate the correspondence between morphological boundaries and those of phonological (e.g. prosodic) categories. In fact, a combination of these two strategies allow Bonet et al. (2007) to formulate an output optimization analysis that captures the seemingly anti-optimizing Haitian Creole pattern in (32). First, the /-a/ allomorph is stipulated to be the default (that is, the morpheme is represented as /{-a > -la}/). As a result, it is effectively only the /-la/ variant whose occurrence (after consonants, (32a)) is seen as being phonologically conditioned; the choice of [pa.pa.-a] over *[pa.pa.-la], or of [pa.pje.j-a] over *[pa.pje.-la], is simply a matter of using the default (highest-priority) allomorph. Second (and following Klein 2003), the selection of /-la/ after consonant-final bases is triggered by a constraint R-Align-Stem-Syll (‘Align the right edge of the stem with the right edge of a syllable’), which forces the stem-final consonant to be parsed as a syllable coda. The output [liv.-la] is thus construed as ‘better’ than *[li.v-a] on the grounds of morpho-prosodic alignment (R-Align), and this is what makes the former emerge as the optimal output.12 As Embick (2010) points out, however, the appeal to phonology here is entirely ad hoc and without independent support from other facts of the language. The constraint R-Align otherwise plays no role in the morphophonology of Haitian Creole. In fact, consonants routinely resyllabify across morpheme boundaries, resulting in rampant R-Align violations in surface forms. While Bonet et al.’s (2007) full analysis does account for this, the larger question that remains is to what extent the analysis as a whole is learnable. The challenging task that a learner has to accomplish is manifold: (i) to infer that the alternation is due to allomorphy, not an active phonological process; (ii) to infer appropriate input representations, one for each allomorph; (iii) to infer an ordering relation (if needed) among the allomorphs; and (iv) to infer a ranking of constraints (consistent with other facts of the language) that accounts for the distribution of the allomorphs. All of these different aspects of the task are mutually dependent on each other (especially (ii)–(iv)), yet all are to be carried out in parallel. Formal learning models that are designed to accomplish this have yet to be developed. (In general, little attention has been paid to learnability concerns in the literature on 12 The alternative *[liv.-a], which uses the default allomorph /-a/ while still respecting R-Align, is ruled out by a constraint *C.V prohibiting CV sequences from being parsed into separate syllables. While this constraint would appear to duplicate the work of Onset, which bans onsetless syllables, the two must be separate and straddle the default-protecting constraint (*C.V ≫ Priority ≫ Onset). Otherwise /-la/ would also be selected after vowel-final bases like /papa/ in order to avoid the onsetless final syllable of [pa.pa.-a]. Bonet et al. (2007) view *C.V as part of a larger family of sonority-based markedness constraints regulating syllable contact.

194

gunnar ólafur hansson

phonologically conditioned allomorphy.) By contrast, in place of (ii)–(iv) a subcategorization learner has only to discover the two frames in (35) or their equivalent, a far simpler and easier task.13 (35) Subcategorization frames for definite marker allomorphy in Haitian Creole a.

[ [ C # ]stem /-la/def sfx ]def word

b.

[ [ V # ]stem /-a/def sfx ]def word

8.5 Conclusions

.............................................................................................................................................................................

This chapter has touched upon a number of ways in which the phonology of a language can interact with aspects of its morphological structure, and different approaches to the modelling of such interactions in a formal-generative framework of grammar were contrasted. We have seen how phonotactic restrictions and phonological processes can be curtailed, triggered, or extended as dictated by particular morphological constructions and constituent-structure configurations. Also examined were some instances in which phonological criteria influence the exponence of morphological properties, such as in the case of phonologically conditioned allomorph selection as well as phonologically motivated paradigm gaps. Exponence in general is, of course, a morphology–phonology interface issue by definition. Since inflectional exponence is allotted its own chapter-length treatment (Trommer and Zimmerman, this volume), I have ignored such important topics here as prosodic morphology (e.g. reduplication, infixation, truncation, templates), featural affixation, mutation, and the like. Another problem not addressed in the present chapter, which is akin to that of allomorph selection (Section 8.4), is the question of phonologically conditioned affix order (Paster 2009). Of particular relevance is so-called ‘mobile affixation’, in which an affix is either prefixed or suffixed depending on phonological characteristics of the base; the best known example is Huave (Noyer 1994; Kim 2010; Trommer and Zimmerman, this volume). The existence of variable affix order governed by phonological well-formedness conditions remains a subject of debate; Paster (2009: 36) sceptically concludes that it ‘does not really exist’, noting that ‘fifteen years after the publication of Noyer (1994), Huave remains the single most convincing example’. Research on morphology–phonology interface issues often raises fundamental questions about the division of labour between phonology and morphology in synchronically accounting for morphophonological alternations and related phenomena. This is particularly true for the interface of phonology with inflectional morphology, given the generally productive character of the latter. Much of the subject matter that forms the 13 The subcategorization frame in (35b) could easily be expanded into separate frames for /-a/, /-ja/, and /-wa/ based on the quality of the vowel. Here it is assumed that the insertion of a homorganic glide [j] or [w] as hiatus-breaker is an easily recoverable aspect of Haitian Creole phonology.

phonology

195

empirical basis of theoretical phonology consists of sound patterns and alternations that, upon closer scrutiny, are limited to some extent by morphological and/or lexical factors (in ways that are often glossed over in analytical works adducing them as examples). From a phenomenological perspective, one might say that there exists a cline from regular and transparently phonological alternations at one end (which it is reasonable to attribute to the phonological grammar) to morphologically stipulated and phonologically suppletive ones at the other (in which the phonological grammar plays no obvious role). Along the middle of this cline fall the kinds of situations that have been the subject of this chapter; for example, ones that seem to involve constructionspecific phonology or phonologically conditioned allomorphy. These in turn range from cases that look suspiciously ‘phonological’ to others for which the involvement of phonological computation and well-formedness considerations seems more remote (e.g. the allomorph selection patterns in Basque and Haitian Creole, respectively). The existence of such a cline (as a purely observational notion) is not surprising, since morphophonological alternations most often originate in low-level, automatic (often allophonic) alternations which are themselves the product of phonetically motivated sound changes (Hansson 2011), but whose defining properties have become partly obscured by subsequent historical changes. The question for phonological and morphological theory is where, and by what criteria, to draw the line demarcating which phenomena involve calling upon the phonological grammar (some version of our function ϕL from Section 8.2 above) and which do not. So long as phonology is modelled as a function relating input (lexical, phonemic) representations to output (surface, phonetic) representations, and so long as at least some alternations in surface shape (allophonic and/or neutralizing) are treated in terms of unique underlying forms undergoing changes in the input–output mapping, such a demarcation line must exist. For the past half-century, the standard heuristic in generative phonology has been to ascribe any alternation to the phonological grammar by default, even at the expense of relatively abstract input representations or complex interactions of rules/constraints; analyses in terms of allomorphy or morphological stipulation are treated as a last resort. Many of the proposals described in this chapter, such as the invoking of constraints on paradigm uniformity or anti-homophony, or the notion of output optimization in allomorph selection, are best understood in this light. Disagreements on the particular theoretical issues are typically played out in the arena of descriptive adequacy and typological coverage (as well as formal parsimony). What has been less often addressed is the question of explanatory adequacy and how the language learner comes to reason her way, on the basis of positive data, to the assumed analysis of the morpho-phonological phenomenon in question. With the increased interest in the formal modelling of phonotactic and morpho-phonological learning in recent years (for an overview, see Albright and Hayes 2011), this is likely to change.

chapter 9 ........................................................................................................

PERIPHRASIS AND INFLECTION ........................................................................................................

andrew spencer and gergana popova

9.1 Introduction

.............................................................................................................................................................................

This chapter is about the relationship between the expression of inflectional information synthetically, by morphology, and periphrastically, or analytically, by syntax. By no means all of the syntax which expresses functional properties should be regarded as periphrasis. Our starting assumption is that there are certain types of syntactic construction which share crucial properties with inflectional morphology, provided we view inflection from the vantage point of paradigm structure. For instance, there are inflectional paradigms which have cells for which there is no synthetic inflected word form, and the content of those cells is expressed by a syntactic construction. However, in some cases that syntactic construction has acquired properties which make it look less like ordinary productive syntax and more like a specially dedicated construction whose role is to fill the cells in that defective paradigm. These are the constructions we refer to as specifically periphrastic and we will identify a number of criteria for distinguishing such constructions from ordinary syntax. In principle, the term ‘periphrastic’ could be applied to lexical forms (such as English particle verbs) or to derivational or valence relations (such as light verb constructions or the English make causative), but most discussion in the literature has focused on inflectional periphrasis (Spencer 2006a; for recent discussion of derivational periphrasis, see Ackerman et al. 2011). In this chapter we will use the term ‘inflection’ to mean ‘synthetically (morphologically) expressed inflection’. We take the term ‘analytic’ to be a synonym of ‘periphrastic’. A periphrastic expression is sometimes called a periphrase. As a first approximation we can say that inflectional periphrasis refers to ‘the expression of inflectional properties by means of a multiword construction’. The multi-word construction generally conforms to some degree to the syntactic norms of the language

198

andrew spencer and gergana popova

(we will return to this later), so that we can think of periphrasis as inflection whose exponent is a syntactic construction. At first blush, this characterization seems to be broadly speaking the way that periphrasis was understood in traditional grammar (Haspelmath 2000 has a useful discussion of the history of usage of the term). For instance, Allen and Greenough’s New Latin Grammar (first published 1888, revised 1903) discussed two Periphrastic Conjugations. The first is the periphrastic future based on the future participle in -urus with forms of the verb esse ‘be’: amaturus sum ‘I am about to love’. The second is a periphrastic construction based on the gerundive in -ndus and the verb esse, denoting ‘obligation, necessity, or propriety’ and having a passive voice interpretation (Greenough et al. 1983: 106–8): amandus sum ‘I am to be, must be, loved’. However, appearances are deceptive. We will claim that the perfective (‘perfectum’) passive construction (amatus sum ‘I was loved, I have been loved’) is a more canonical type of periphrasis, yet this construction is not even described as a periphrase in the traditional grammar of Allen and Greenough. All aspects of our characterization of periphrasis, ‘inflectional property’, ‘word’, ‘multi-’, and ‘construction’, require clarification, and we will be discussing some of them in detail. Here we will briefly sketch these problems. Our most problematical notions are ‘word’ and hence ‘multi-word’: to identify a periphrasis we need to individuate words in the construction and distinguish them from loosely cohering affixes and clitics. There are well-known difficulties here, of course. We will therefore try to restrict ourselves to relatively uncontroversial cases, though we will be illustrating some aspects of this problem in our discussion of Bulgarian periphrases in Sections 9.8 and 9.9. We will take ‘inflection’ to mean an obligatory (hence, productive) functional (grammatical) property associated with a class of lexemes and expressed by means of morphology. Inflection therefore defines the set of inflected word forms of a lexeme realizing the inflectional properties of that class of lexemes. By contrast, derivational morphology defines relations between a base lexeme and a distinct, derived lexeme, with a semantic representation which subsumes that of the base lexeme. For those who reject the lexeme concept and realizational models, inflection is simply morphology which expresses a functional property. However, inflectional paradigms are sometimes defective, in which case inflection is not obligatory (Baerman this volume; Hansson this volume). Conversely, a derivational process may be highly productive with few, if any, defective lexemes and thus, in effect, obligatory. Moreover, it may add content which is highly abstract and hence similar to that of a functional, grammatical element. English subject nominalizations such as driver from drive approach this situation, as do verbs prefixed with re- (which behaves to some extent in the way that an iterative Aktionsart marker might behave in other languages). Periphrases have been approached broadly from two different perspectives: researchers have looked at their place in morphological paradigms on the one hand, and have looked at the syntactic properties of the constructions that constitute the periphrases on the other. These two different perspectives are often, though not necessarily, linked. We will try to keep them apart for the purposes of the discussion to

periphrasis and inflection

199

follow, but will note the links between them when appropriate. Of these two perspectives the relationship of periphrasis to paradigms is the more important one. Indeed, this relationship is often taken to be definitional. For example Haspelmath (2000: 655) defines periphrasis as ‘a situation in which a multi-word expression is used in place of a single word in an inflectional paradigm’, and Brown et al. (2012) take this to be the canonical type of periphrasis. We will discuss the place of periphrasis in the morphological paradigm in Section 9.3.1, but before that we briefly survey the types of constructions that have been described as periphrastic.

9.2 Types of Periphrastic Construction

.............................................................................................................................................................................

We briefly sketch some of the properties most commonly expressed periphrastically. We outline first the verb properties, which provide the commonest forms of periphrasis, and then look briefly at nominal properties (of nouns and adjectives). Nominal periphrases are much less commonly discussed than verbal periphrases and our discussion will be somewhat speculative.

9.2.1 Verbs Tense/aspect/mood categories are very commonly expressed periphrastically. Negation is often expressed periphrastically by means of an inflecting auxiliary (Spencer 2012b). Voice categories such as passive and causative are also frequently periphrastic (see Ackerman and Webelhuth 1998 for extensive exemplification). In addition, following Ackerman and Webelhuth (1998) we could add the notion of ‘predication’ here. In many languages predicative noun or predicative adjective constructions are expressed synthetically: the noun or adjective is inflected to agree with the subject and in some cases inflects for tense, for instance, in Nenets (Salminen 1998: 539). We will discuss the tense/aspect and negation system of Bulgarian in more detail in Sections 9.8 and 9.9. One of the problems facing the descriptive grammarian is distinguishing between periphrases (with auxiliary verbs, say) and other types of verb construction, particularly serial verbs and light verb constructions, which are better thought of as expanding the lexical inventory of the language or as special syntactic constructions in their own right.

9.2.2 Nouns Among the nominal properties, the candidates for periphrastic expression include number, case, definiteness, and possession.

200

andrew spencer and gergana popova

Tagalog has no morphological category of nominal number and bare nouns are vague as to singular or plural reference. However, a plural interpretation can be made explicit by means of the proclitic mga (pronounced /maŋah/) (Schachter and Otanes 1972: 111–13). This can occur either before the noun head or at the left edge of the nominal phrase (Schachter and Otanes 1972: 112) (1)

a. mga para sa batang libro plural for the child book b. para sa batang mga libro for the child pl book ‘books for the child’

(Tagalog)

However, it is not clear that we should regard such constructions as periphrastic. We could equally treat the plural marker as some kind of determiner or quantificational modifier, along the lines of English many. Brown et al. (2012) cite a more convincing instance of a periphrase with the oblique case forms of the dual number paradigms of the Samoyedic language Nenets. Nouns inflect for number (singular, dual, plural) and for seven cases, but in the dual number the four local cases (dative, locative, ablative, prosecutive) are expressed by means of the bare noun combined with the casemarked forms of the postposition nya- ‘at’ (Salminen 1998: 537). Since other cases are expressed synthetically in the dual number (and in singular and plural) we can regard the periphrasis as part of the morphological inflectional paradigm (in other words, the Nenets case is a clear instance of what we will later call an intersective periphrasis; see Section 9.3.1). Many languages express definiteness morphologically, while others express definiteness by means of a separate function word (definite article/determiner). Danish has an indefinite article for singular nouns, en, common (C) gender, et, neuter (N) gender, and a suffixal definite article -en/et (C/N):

(2) Danish (unmodified) definite nouns singular indef def en mand mand-en et barn barn-et

plural indef def mænd mænd-ene børn børn-ene

‘man[C]’ ‘child[N]’

However, when the noun is modified by an adjective that noun no longer takes the definite suffix. Instead the whole phrase is marked by a phrase-initial definite determiner den/det/de:

periphrasis and inflection (3)

201

Danish modified definite nouns singular den unge mand ‘the young man’

plural de unge mænd ‘the young men’

det glade barn ‘the happy child’

de glade børn ‘the happy children’

One might argue that this is an unusual instance of feature intersection, in which a morphologically expressed category is expressed periphrastically in specific syntactic contexts (rather than specifically morphological contexts). One might also argue that the indefinite article is an instance of intersective periphrasis, given the morphologically-expressed definite forms. Many languages distinguish alienable from inalienable possession. It is not uncommon for such languages to express inalienable possession morphologically but alienable possession by means of some kind of preposition or some other type of function word. In Oceanic languages, for instance, inalienably possessed nouns are typically inflected with pronominal affixes cross-referencing the possessor, while alienable possession is typically expressed by a distinct function word, a ‘possessive classifier’, which cross-references the possessor, usually with the same possessive morphology found with inalienably possessed nouns. Lichtenberk (2009) surveys the construction and introduces the basic phenomenon with examples from Lolovoli (Lichtenberk 2009: 253):1 (4)

Lolovoli possessive constructions a. hava-da family-nonsg(incl.):poss ‘our family’ b. no-da hala poss.clf-nonsg(incl):poss visitor ‘our visitor’

inalienable

alienable

Lichtenberk (2009: 273–6) points out that the possessive constructions in Oceanic languages often exhibit ‘fluidity’, by which he means that one and the same noun may receive different possessive marking depending on the semantics/pragmatics of the expression. In Manam, for instance, an article of clothing is treated as inalienable (a little like a body part) when it is being worn, but as alienable otherwise: (5)

1

Manam possessive constructions a. baligo-gu grass.skirt-1sg:poss ‘my grass-skirt (when I am wearing it)’

inalienable

Abbreviations follow Lichtenberk’s usage: incl = ‘inclusive’, poss = ‘possessor’, clf = ‘classifier’.

202

andrew spencer and gergana popova b. baligo ne-gu grass.skirt poss.clf-1sg:poss ‘my grass-skirt (when I am not wearing it)’

alienable

Assuming that the possessive classifiers are treated as autonomous syntactic words in the grammars of these languages, we could argue that the possessor construction paradigm is governed by person/number properties together with a property, say, possessor-type, such that the complete paradigm of possessive forms is defined by combinations of the properties [possessor-type:{alienable, inalienable}, possessor:{1sg, . . . , 3pl}]. Since the part of the paradigm defined by the property set [possessor-type:inalienable, possessor:1sg, . . . , 3pl] is expressed morphologically, we would expect the other part of the paradigm to be expressed morphologically, but this expectation is confounded, and so we have a (minimally) intersective periphrasis.2

9.2.3 Adjectives Adjectives often borrow their morphosyntax from the categories to which they are most closely related in a given language, that is, either nouns or verbs, but sometimes adjectives are claimed to exhibit their own periphrases. A frequently cited instance of periphrastic expression in adjectives is the English comparative/superlative construction, although as we point out in Section 9.3.2 it is not clear to what extent this can really be called a periphrasis.

9.3 Periphrasis and Paradigms

.............................................................................................................................................................................

9.3.1 The Canonical Type—Intersective Periphrasis Synthetic inflected forms in a language are generally organized in paradigms. The notion of paradigm is often understood simply to be the set of inflected word forms of lexemes (see Blevins, this volume, for discussion). If we take this unstructured view of paradigms then periphrasis need not exist as a theoretical notion, at least not on the definition assumed by Haspelmath (2000) in Section 9.1. It is a characteristic of much 2 One might argue that English represents a similar pattern. It has been argued that the English possessive -’s is ambiguous between a ‘head genitive’ and a ‘phrasal genitive’ function (Payne and Huddleston 2002: 479; Payne 2009). In an expression such as Edward’s daughter or somebody’s initiative the -’s element is said to be a genitive case marker. There are two arguments for this: (i) the head genitive of the 1sg pronoun is my, while the genitive of a phrase ending in a 1sg object pronoun, say, will end in the form me’s: my hat vs. the man opposite me’s hat; (ii) the -’s element exhibits haplology when combined with a word ending in plural -s: the ducks’ plumage vs. ∗ the ducks’s plumage. This haplology is taken as evidence that the -’s formative is an affix, not a clitic, and hence that the forms are inflected with a genitive suffix. But in that case we have a morphological construction which alternates with a syntactic construction (under conditions which are far from well understood, see, for instance, Rosenbach 2002), and hence, apparently, the of -genitive is periphrastic. Needless to say, this is a marginal instance of periphrasis, if indeed it can be viewed as periphrasis at all.

periphrasis and inflection

203

recent work on inflection that paradigms are structured; however, the most salient form of structuring is that implied by intersecting feature systems, in which single forms express or realize values for two or more features simultaneously. For example, a typical Latin verb form will express simultaneously tense/aspect/mood, person and number, and voice, as illustrated in (6): (6)

Latin conjugation (present indicative)

1 2 3

active singular plural amo amamus amas amatis amat amant

passive singular plural amor amamur amaris amamini amatur amantur

A feature system of the kind familiar from most approaches to morphology automatically defines an (abstract) paradigm, namely the space of all the logically possible combinations of feature values. In general, a grammar has the option of reducing this combinatory space by means of feature co-occurrence restrictions. For example, a Latin verb in the subjunctive mood can inflect only for past and present tense, not for future tense. Any permissible combination of feature values defines a cell in the paradigm. Other things being equal, such a cell must be associated with a form of the lexeme, the inflected word form that is said to occupy that cell. If, for some reason, no such synthetic form exists, we potentially have a ‘gap’ in the paradigm . If the gap remains we have a defective paradigm (Baerman, this volume). However, in some cases there is no gap because the cell in question is filled by a periphrastic construction. For instance, the perfective passive in Latin, illustrated in (7), expresses the properties perfective aspect, passive voice in addition to person/number properties. In the active voice, both imperfective and perfective aspects are expressed synthetically, while in the passive voice it is only the imperfective aspect that is expressed synthetically. The combination of the two properties passive voice and perfective aspect has to be expressed periphrastically. (7)

Latin conjugation—active/passive active singular plural Imperfective 1 amo amamus 2 amas amatis 3 amat amant

amor amaris amatur

amamur amamini amantur

Perfective 1 amavi 2 amavisti 3 amavit

amatus sum amatus es amatus est

amati sumus amati estis amati sunt

amavimus amavistis amaverunt

singular

passive plural

204

andrew spencer and gergana popova

We will refer to this kind of periphrasis as ‘intersective’. The feature combinations that give rise to it are the properties of aspect and voice. In the general case, these can be combined with each other to give cells that are realized synthetically. The combination perfective passive is therefore anomalous with respect to the other logically possible combinations. We take the intersective type of periphrasis to be the most clear-cut, indeed, the canonical type, because the relation between the syntactic expression and the structure of the morphological paradigm is evident. The role of paradigm structure can help us to weed out false instances of periphrasis. The English modal auxiliaries, such as can, are defective and have only a present tense form (and, perhaps, a preterite form could). However, it is possible to express infinitive and perfect aspect modalities by means of a construction with the adjective able: to be able, has/had been able. As Börjars et al. (1997) point out, however, this is not (suppletive) periphrasis. The crucial point is that the construction (be) able has a full paradigm of its own. Since its meaning covers some of the meanings of can, it turns out that some forms of be able accidentally correspond to the missing forms of can. In other words, a periphrasis is a bona fide element in the paradigm of a lexeme. It is not merely a syntactic expression which just happens to express the meaning of the missing synthetic form.3 Haspelmath (2000) calls our intersective type of periphrasis ‘suppletive type (a)’ and points out that it helps achieve paradigm symmetry. We do not adopt the term ‘suppletive’ for intersective periphrasis because it has unwelcome connotations of irregularity.4 There are, of course, paradigms in which a cell is occupied by a suppletive or partially suppletive irregular form but this phenomenon, like defectiveness, must be carefully distinguished conceptually from intersective periphrasis, which is a rule-governed phenomenon.

9.3.2 Non-canonical Cases of Intersective Periphrasis We have said that the least controversial instances of periphrasis are those with the intersective property. However, applying the intersective criterion is sometimes not entirely clear-cut. There always remains the possibility that we have a defective paradigm and that there exists a parallel syntactic construction with roughly the same meaning that tends to get used in those contexts where we would have used the nonexistent morphological form, as in the case of the ‘inflected’ forms of can discussed above. We begin with a periphrastic construction found in Classical Greek conjugation which closely parallels the Latin perfect passive periphrasis (see also Haspelmath 2000: 658). Greek verbs distinguish a variety of tenses, including a perfect series (present and 3 That is essentially the account of the Latin passive perfect based on the notion of blocking proposed by Kiparsky (2005). 4 The relation between suppletion and periphrasis is explored in Vincent and Börjars (1996).

periphrasis and inflection

205

past) and three voices (active, middle, and passive. The middle and passive are largely identical, and often called the ‘medio-passive’). The medio-passive of luo ‘I free’ and grapho ‘I write’ are shown in (8) (there is no first person dual form in Greek):5 (8)

Classical Greek medio-passive: lelumai ‘I have freed (myself, for myself)’, gegrammai ‘I have written (for myself)’ 1sg 2sg 3sg 2du 3du 1pl 2pl 3pl

lelumai lelusai lelutai lelusthon lelusthon lelumetha lelusthe leluntai

gegrammai gegrapsai gegraptai gegraphthon gegraphthon gegrammetha gegraphthe gegrammenoi eisi

In general the 3pl form of any verb in any tense/aspect/mood/voice form is expressed synthetically, as with the perfect middle of luo. But where the verb stem ends in consonant, as in the case of grapho, the 3pl form is expressed periphrastically, using the participle and the 3pl present tense form of the verb ‘to be’ (in the pluperfect the past tense of the auxiliary is used). It is only in the perfect series of the medio-passive that we find this substitution. The other forms are synthetic: active perfect gegraphasi ‘they have written’, future perfect passive graphe:sontai ‘they will have been written’, and so on. Hence, we must conclude that the periphrastic construction occupies a cell in what ought, by rights, to be a purely morphological paradigm. In the case of the consonant-stem perfect middle forms, we have a very specific case of periphrasis, occasioned presumably by morphophonology (adding the expected person/number ending -ntai to the consonant stem would result in an unacceptable cluster). It might be argued then that the periphrasis is simply a resort to a synonymous syntactic construction in the face of impossible morphology. However, although this might provide a diachronic, etymological explanation for the situation it does not alter the fact that we are dealing with an instance of intersective periphrasis (note that the language could equally well have left the paradigm defective).6 Elsewhere in the conjugation system of Greek we find a more systematic instance of periphrasis (illustrated in the tables in Weir Smyth 1920: 115), comparable to 5

We transcribe aspirated consonants as digraphs ‘ph’, ‘th’, and upsilon as ‘u’. Matters are slightly more complex than this, because the periphrasis, apparently, was also sometimes used for other person/number combinations, where there is no morphophonological motivation. Thus, Weir Smyth (1920: 183) writes ‘[t]he perfect or pluperfect passive is often paraphrased by the perfect participle and estí or ê; gegramménon estí ‘it stands written’. (The forms estí, ê are third person singular forms of ‘be’ in the present and past tense respectively). So to some extent, then, we may be dealing with overlapping paradigms rather than a pure form of intersective periphrasis. 6

206

andrew spencer and gergana popova

the Latin passive perfect case. In the active voice the perfect indicative tense and the subjunctive, optative, and imperative moods are generally expressed synthetically, for all verbs: leluka ‘1sg perfect indicative’, lu:so:/lu:saimi ‘1sg subjunctive/optative’, lu:son ‘2sg imperative’ (from luo). However, the combination of perfect tense and subjunctive/optative/imperative mood is systematically expressed periphrastically for all person/number combinations (using a participle inflected for number/gender and the auxiliary be): leluko:s o:/e:se: ‘perfect subjunctive, 1sg/2sg/3sg’, lelukotes o:men/e:te/e: ‘perfect subjunctive, 1pl/2pl/3pl’, and likewise leluko:s eie:n . . . ‘perfect optative 1sg, . . . ’, leluko:s isthi . . . ‘perfect imperative 2sg . . . ’. Here we have periphrasis which is not phonologically, lexically, or semantically conditioned. Haspelmath also discusses the periphrastic comparative/superlative forms of English adjectives. Monosyllabic or disyllabic adjectives ending in an unstressed syllable have synthetic forms in -er/est, but other adjectives have to combine with the adverbs more/most: ∗ beautifuller vs. more beautiful. He calls this ‘suppletive periphrasis type (b)’, whose function is not to fill in the gaps in the paradigm across a lexical class (for example, all Latin verbs), but rather to fill in gaps in the paradigms of a subclass (for example, adjectives in English of more than two syllables). The difficulty with this example is that monosyllabic adjectives that have the suffixed forms can also often have the ‘periphrastic’ forms, although acceptability often depends on somewhat complex factors such as word frequency. For instance, although it is very difficult to accept more good in the face of better, for us, examples such as (much) more green, (a little) more dear, (slightly) more happy are no less acceptable than greener, dearer, happier. But even more good is acceptable in certain contexts: She wasn’t noticeably more good to her own children than to other people’s, where good takes the complement to her own children, and the comparative more itself is modified. What this suggests is that all adjectives (broadly speaking) can take more/most but that this option is dispreferred for very common adjectives with standard interpretations. In that case we have a synthetic comparative/superlative form which overlaps with the more general syntactic (analytic) construction. For such examples, therefore, we cannot really speak of an intersective periphrasis (although we might call it a periphrasis on other grounds). Haspelmath (Haspelmath 2000: 657–9) points out other problems in pinning down instances of intersective periphrasis. One of these is found when the inflectional forms themselves fail to define a clear pattern, so that it is difficult to judge whether the analytic construction occupies an inflectional cell or not. One spectacular instance is that of the Hungarian ‘future tense’ construction. The verb lenni ‘to be(come)’ is unique in having a synthetic future tense. With all other verbs the auxiliary fogni is used to express future time (although whether we would really want to call this a periphrastic tense form is a different matter). Rather than saying that all verbs except lenni have a periphrastic future, we can follow Haspelmath in saying that lenni has an anti-periphrastic future. Similarly, Russian imperfective verbs form their future by combining the infinitive with the future copular/auxiliary form, budu, budeš, budet, . . . : budu spat ′ ‘I will sleep’. The copula/auxiliary byt’ itself is unique in being the only imperfective verb with its own

periphrasis and inflection

207

future form: budu na rabote ‘I will be at work’ (∗ budu byt’ na rabote).7 Other examples that can be cited are those languages in which some, but not all, adpositions inflect for the person/number of their complements. This is found with Celtic prepositions, for instance, in Welsh (Fife and King 1998: 483–5); (Williams 1980: 127–30), where most of the more frequent prepositions have ‘conjugating’ forms. The preposition gan ‘with’ has the inflectional paradigm shown in (9): (9)

Inflecting preposition gan ‘with’ (Welsh) singular plural 1 gennyf gennym 2 gennyt gennych 3m ganddo (ef) ganddynt 3f ganddi (hi)

(The 3sg forms in parentheses are accompanying clitic pronominals. There is no gender distinction except with 3sg forms.) Williams lists fifteen such prepositions, inflecting with varying degrees of irregularity. On the other hand, gyda, ‘with’, has no inflecting forms, and nor do prepositions such as efo, ‘with’, erbyn, ‘against’, herwydd, ‘according to’, hyd, ‘as far as’, nes, ‘until’, wedi, ‘after’, and a number of others. Most of these combine with full form pronominal complements: gyda fi/ti/fe ‘with me/thee/him’. The inflected forms are obligatory with pronominal complements. This leaves open the possibility that the analytic forms are intersective periphrases. However, it is unclear to what extent we would want to maintain that antiperiphrasis was a distinct phenomenon in its own right. We often find that (usually highly frequent) lexemes realize in their morphological paradigms a greater number of morphosyntactic properties and contrasts than other lexemes. The verb be in English is a well-known example, distinguishing a unique 1sg present tense form am and a unique singular/plural contrast in the past (was/were). In the case of the overdifferentiated past tense forms there is no anti-periphrasis because all other verbs simply fail to make any such contrast. On the other hand, it would be open to us to argue that am is an anti-periphrastic form, and that 1sg is usually expressed periphrastically with a 1sg pronoun (as, indeed, in the past tense of ‘be’, I was). The ‘anti-periphrasis’ cases are then simply cases of overdifferentiation in the morphological paradigm for properties that are normally expressed syntactically. They therefore relate to examples of portmanteau forms discussed below.

9.4 Periphrasis and Inflection

.............................................................................................................................................................................

Since periphrases realize the content of a cell in an inflectional paradigm they share with morphological inflection the property of being obligatory (see Section 9.1). Thus, 7

Interestingly, the iterative form of byt’, byvat’, forms its future regularly: Budu byvat’ na rabote po ˇcašˇce ‘I.will be at work more often’, i.e. ‘I will come to my workplace more often’.

208

andrew spencer and gergana popova

all transitive Latin verbs can be expected to share the perfective passive periphrasis. Indeed, Latin also has deponent verbs, which are active in meaning but passive in form. The perfective active paradigm of such verbs is formed periphrastically, therefore. Some verbs in Latin are defective and have only present tense series forms (e.g. ferio ‘strike’, Greenough et al. 1983: 117). Such verbs therefore lack periphrastic perfective passives. As far as we can tell there are no Latin verbs which are defective specifically with respect to periphrastic forms. This seems to be a corollary of a more general property: periphrases show none of the irregularities associated with morphologically expressed forms (though they may certainly show the kinds of irregularity associated with syntactic constructions, such as having idiomatic interpretations). For instance, if periphrasis can express the same relations as morphological inflection then we might expect to see instances of defectiveness, where some lexemes fail to enter into a particular periphrastic construction or part of a periphrastic paradigm. Assuming that, say, the English aspectual/voice auxiliary constructions are periphrastic, a hypothetical example of such defectiveness would be a dynamic verb lexeme, say, schmopen with the same meaning as open but from which the perfect cannot be formed: he is schmopening it, it was schmopened (by him), but ∗ he has schmopened it. Inflectional defectiveness is widespread (see, for instance, the papers in (Baerman et al. 2007), and seems to be found in pretty well all inflectional systems, even in highly reduced morphological systems such as that of English. However, we have not found any such examples of defectiveness in periphrasis. If periphrasis shared all the properties of morphological expression we might also expect to see the periphrastic equivalent of uninflectable lexemes, such as English sheep, which appear in appropriate morphosyntactic contexts but as bare forms of the lexeme. Thus, our hypothetical schmopen as a periphrastically uninflectable lexeme would have the forms he schmopen, for the progressive and perfect and it schmopen (by him) for the passive. It is particularly difficult to imagine instances of defectivity and uninflectability in periphrasis, presumably because the exponence is effectively a syntactic construction and these are less prone to such irregularities. It is therefore much more difficult to imagine a plausible grammaticalization path that would lead to such defectiveness in periphrasis. The Latin perfective passive intersective periphrasis represents for us a clear example in which the analytic form is integrated into the inflectional paradigm. However, there are many cases in which it is less clear how the inflectional paradigmatic space is to be defined. One common difficulty is that of distinguishing between a true affixal morphological system and a clitic system, because clitics assume increasingly affixlike properties with grammaticalization (Spencer 2006a; Spencer and Luís 2012). We discuss this problem further in Section 9.8. A broader view of periphrasis would have the term apply to any construction that expresses a grammatical feature, even if there are no values of this feature that are expressed morphologically. Haspelmath (2000: 656) calls this type of periphrasis categorial. If we extend the notion of periphrasis in this way, we may wish to define it

periphrasis and inflection

209

slightly differently from the definition offered in Section 9.1. On this understanding ‘[a] periphrastic expression is simply one which expresses a grammatical meaning in a multi-word construction’ (Haspelmath 2000: 660). For instance, we might claim that the English progressive and perfect aspects are periphrastic, even though the property aspect is never expressed synthetically in English. Similarly, we could claim that the English comparative/superlative in more/most was an instance of categorial periphrasis. This more general definition would make identifying periphrasis less bound with the facts of a particular language, as it makes it independent of paradigms. Indeed, on this definition any construction containing function words would be periphrastic, even in a language which lacked any morphology proper whatsoever (Haspelmath 2000: 655) . Such a wide characterization would give rise to counterintuitive analyses. Consider, for example, English yes–no questions. These are expressed by inverting the first auxiliary with the subject. When the corresponding uninverted declarative form lacks an auxiliary verb the dummy auxiliary do is introduced. But, on the liberal interpretation of periphrasis, this would mean that the English interrogative mood is expressed syntactically, by pure word order alternation, for modal auxiliary constructions and progressive, perfect, and passive, but periphrastically for active voice, simple aspect. This is because it is only the active voice, simple aspect forms which, apparently, are realized by means of a function word. This is a very counter-intuitive analysis and one which seems entirely arbitrary. We must therefore ensure either that our characterization of periphrasis fails to entail such a conclusion, or that the conclusion is carefully justified.

9.5 Periphrasis and Syntactic Structure

.............................................................................................................................................................................

As we said above, there are (at least) two general ways of looking at periphrasis. The first one is to look for the place of periphrasis in the paradigmatic structure of a language. The other one is to look for unique characteristics of the multi-word expressions one wishes to call periphrastic. For some scholars it is the nature of the construction that is the crucial property in characterizing it as a periphrasis. Discussions here often link periphrastic constructions to grammaticalization, but pay specific attention to its syntactic and semantic effects. In particular, we often see analyses which postulate the transformation of lexical elements into functional ones with concommitant semantic bleaching and/or semantic non-compositionality. For example Ozonova (2006: 9) refers to definitions of periphrases (what she calls analytic constructions) as a combination of a main (lexical) and an auxiliary element, with the main elements defining the lexical meaning of the whole construction and the auxiliary element defining the grammatical meaning associated with it. Functionally and semantically a periphrase behaves as a single word, but structurally it is a multi-word construction. Another assumption underlying the

210

andrew spencer and gergana popova

tradition summed up by Ozonova is that only one of the elements of a periphrase will inflect (the auxiliary element that carries grammatical meaning), whereas the other will be an uninflected form. (Ozonova is careful to point out that we cannot attribute the grammatical meaning of the whole entirely to the auxiliary element—see our discussion of non-compositionality later in this section). A similar view of periphrasis is expressed by Žirmunskij (1976: 87). He provides an explicit link between the nature of the construction and its integration into the inflectional system of a language by highlighting the functional similarities between auxiliary words and affixes: ‘Since it forms part of an analytic [i.e. periphrastic] construction the function word is grammatically equivalent to an inflectional morpheme (affix). At the same time, the analytic construction is a form of the corresponding content word, and enters into its inflectional system, the word’s so-called “paradigm”’ (emphasis original).8 The cases of non-intersective periphrasis which seem most convincing to us are indeed those in which the function word component itself inflects for some of the properties expressed by the construction as a whole, as in the case of the English perfect/progressive aspects. Such a treatment seems especially attractive if the function word realizes properties which would normally be expressed morphologically on the lexical verb in other parts of the paradigm. A case in point is the negation construction in the Southern Tungusic language Udihe (Nikolaeva and Tolskaya 2001: 211, 214, 818), in which a fully conjugating negative auxiliary combines with an uninflected form of the lexical verb. Since the negative auxiliary word is the only exponent of polarity we cannot say that there is feature intersection for the property [Polarity:{affirmative, negative}]. We should note that if we admit constructions such as Tungusic negation as a type of periphrasis, we would be forced to treat as periphrastic a number of constructions which are not usually so regarded. For instance, in Spoken French there is a singular/plural distinction but only a small number of irregular nouns actually have a distinct plural form. It is the definite and indefinite determiners that obligatorily signal the number of the noun: la table /latabl/ ‘the table’, une table /yntabl/ ‘a table’, vs. les tables /letabl/ ‘the tables’, des tables /detabl/ ‘(some) tables’. Similarly, in Spoken German the four-way case system is reflected principally in the determiner system, since nouns have lost nearly all their case inflections. However, grammars do not speak of periphrastic number in French or periphrastic case inflection in German. The reason is no doubt that the ‘proper’ function of the determiners is not the expression of number/case and so it is seen as accidental that the determiners have become the principal exponent of those properties for most lexemes. But there are no particular grounds for 8 ‘Buduˇci cˇ ast’ju analitiˇceskoj konstrukcii, služebnoe slovo po svoej grammatiˇceskoj funkcii ekvivalentno slovoizmenitel’noj (formoobrazujušˇcej) morfeme (affiksu). Tem samym analitiˇceskaja konstrukcija javljaetsja f o r m o j sootvetstvujušˇcego z n a m e n a t e l ’ n o g o s l o v a, vxodja v sistemu ego slovoizmenenija (ili formoobrazovanija)—v tak nazyvaemuju “paradigmu” slova.’

periphrasis and inflection

211

restricting periphrastic expression to those properties which appear to be realized by function words serving as ‘principal exponents’ rather than ‘secondary exponents’. We mentioned at the beginning of the chapter that periphrases are often syntactically unremarkable, in that they share the properties of other syntactic constructions in a language. Sometimes, however, periphrases are associated with structural peculiarities that set them apart. One such peculiarity could be morphosyntactic non-compositionality, that is, the construction as a whole might be associated with morphosyntactic properties that do not arise from any of the component parts. We return to this property directly. Non-compositional syntax of any kind is noncanonical by definition (although non-compositionality may well be a prototypical property of periphrases—note that prototypical properties of a construction are often non-canonical properties). Since periphrases look more like syntax than morphology the temptation is therefore to (re-)analyse all putative periphrases as instances of syntax. For instance, Kiparsky’s (2005) analysis of the Latin perfect passive periphrasis takes the periphrasis to be the by-product of combining a perfect passive participle (an adjectival form expressing a perfect passive meaning) with the copular verb. Syntactic analyses of this sort will sometimes provide an adequate treatment of periphrastic constructions, but there is one type of situation that they describe rather badly. This is the situation in which the periphrasis is not only intersective but also involves some kind of non-compositionality. For instance, the Latin perfect passive periphrasis admits five tense/mood forms: the future indicative and present and past tense forms of the indicative and subjunctive. The standard way of expressing the past perfect (or pluperfect) passive is to use the imperfective past form of the verb ‘be’: amatus eram/essem for 1sg pluperfect indicative/subjunctive. Now, occasionally we find the perfective past form of ‘be’ used in this construction: amatus fui/fuerim, but the construction with the imperfective forms is sanctioned in the handbooks as the default construction and it is this that is the most frequent textually.9 Whether and to what extent the perfective form of the copula was used is immaterial, however, the point is that it was common to use an imperfective aspect form to construct a perfective aspect periphrase. Thus, there is a clash of aspect features in the common form of the construction. No purely syntactic analysis of any sort seems viable for describing such a situation:10 the cells in the inflectional paradigm of a verb lexeme require (at least as their first preference) a non-compositional construction in order to express the content of those cells. Therefore, the periphrasis has to be the result of some kind of rule of exponence. Indeed, the 9

A brief Google search for interfectus erat/fuit ‘had been killed’ and locutus erat/fuit ‘had said, spoken’ revealed that the construction with fuit is found most often with sources written in medieval times or in modern texts, e.g. from the Vatican, while the ‘standard’ construction with erat, with perfective meaning, is the one favoured in ‘Golden’ and ‘Silver’ period authors, Caesar, Cicero, and Tacitus. 10 It seems to us that such a construction would be particularly difficult to describe in Optimality Theoretic terms, given the existence, and occasional use, of a compositional alternative to the noncompositional variant.

212

andrew spencer and gergana popova

Latin periphrasis is rather strong evidence in favour of realizational models of inflection generally, given that any syntactic treatment would be obliged to at least favour the periphrasis with fui/fuerim over that with eram/essem, if it permitted the latter at all. But non-compositionality of this sort is rather typical of a morphological paradigm, and has been adduced as a criterion for periphrasis (Section 9.6).

9.6 Three Criteria for Periphrasis (Ackerman and Stump 2004) ............................................................................................................................................................................. An important recent contribution to the discussion of the role and status of periphrasis is that of Ackerman and Stump (2004) (based in part on the work of Ackerman and Webelhuth 1998). They explicitly couch their analysis within the inferentialrealizational framework of Paradigm Function Morphology (Stump 2001b: 2002: 2006), one of the class of models built around the principle of realization-based lexicalism (Blevins 2001).11 They argue that we can call a construction a periphrase if it respects any one of three criteria (each one of the criteria is sufficient, though not necessary): Criterion I: If an analytic combination C has a featurally intersective distribution then C is a periphrase (p. 128). Criterion II: If the morphosyntactic property set associated with an analytic combination C is not the composition of the property sets associated with its parts, then C is a periphrase (p. 142). Criterion III: If the morphosyntactic property set associated with an analytic combination C has its exponents distributed among C’s parts, then C is a periphrase (Ackerman and Stump 2004: 147). Criterion I is the intersective criterion already discussed. The other two criteria refer not to the paradigmatic organization of a language, but to the nature of the multiword expression under discussion. Criterion II adduces idiosyncrasy or idiomatic interpretation as an indication that we are dealing with a periphrase. Ackerman and Stump (2004) cite the case of the negative second past in the verb paradigm of the Uralic language, Eastern Mari. The negative second past is expressed by a form they call the affirmative gerundial stem of the verb lexeme combined with a negative present tense form of the copula/auxiliary. Since the auxiliary forms are present tense forms and the lexical verb is in a non-finite form, the property second past tense is not provided by either part 11 See also Ackerman et al. (2011) for a recent discussion of periphrasis within the model of realization-based lexicalism.

periphrasis and inflection

213

of the periphrase. Therefore, we must conclude that the construction is semantically (featurally) idiomatic. The Eastern Mari case is rather similar to the past tense construction in Spoken French: a present tense auxiliary (avoir ‘have’ or eˆ tre ‘be’ depending on the lexical verb12 ) combines with a non-finite perfect participle: Marie a chanté ‘Marie has sung/sang’. In French, the construction can also be interpreted as a present perfect form (which one might wish to argue is a compositional construction). However, the construction is ambiguous: as a perfect aspect form it is in opposition to the past perfect, which is formed from the past imperfect form of the auxiliary (avait): Marie avait chanté ‘Marie had sung’, but as a past tense form it is in opposition to the imperfect past form: Marie chantait ‘Marie was singing’. Moreover, the past perfect can also be expressed by taking the periphrastic form of the past tense of the auxiliary a eu, literally ‘has had’: Marie a eu chanté ‘Marie had sung’. The French and Eastern Mari constructions illustrate what seems to be the most common way in which a periphrastic construction can be idiomatic or noncompositional: the construction as a whole expresses a value of a grammatical property, but some part of the construction expresses a conflicting value of that property. We must exercise caution here, though. Many word forms are underdetermined with respect to a crucial property. For instance, we often find finite verb constructions expressed by a finite auxiliary in combination with a non-finite form of the lexical verb but it would be perverse to argue that this represents non-compositional expression on account of a clash between finite and non-finite property values. Equally, if a language expresses negation by placing a negation particle somewhere in the affirmative form of a clause we cannot argue that we have a feature clash for the value of the property polarity. Criterion III is less obvious than the previous two and we therefore reproduce Ackerman and Stump’s own example by way of illustration. They cite the case of the negative future forms of another Uralic language, Udmurt. This is formed by means of a negative finite auxiliary (ug, ud, . . . ) and a connegative form of the verb that inflects solely for the number of the subject (and then optionally), as illustrated in (10) for the verb mɪ̈nɪ̈ ‘go’: (10)

Udmurt negative future paradigm for mïnï ‘go’ 1 2 3

singular ug mïnï ud mïnï uz mïnï

plural um mïnï(le) ud mïnï(le) uz mïnï(le)

As can be seen, the second/third person forms of the auxiliary are identical in the singular and plural so that the only indication of number comes from the -le inflection on 12

As Matthew Baerman points out to us this seems to be an instance of periphrasis falling into inflectional classes, which is generally a morphological property.

214

andrew spencer and gergana popova

the converb. This is what Ackerman and Stump mean by ‘distributed exponence’: person is marked on the auxiliary while number is marked the connegative forms of the lexical verb, so that person and number marking are distributed across auxiliary and connegative form. This kind of ‘economy of expression’ is said to be typical of morphological systems but not, for instance, typical of those syntactic systems that have abundant agreement processes. It seems that Ackerman and Stump intend that the features that are distributed should be features that we would otherwise expect to go together, by being realized on a single word form, such as pronominal person/number features on verbs, which, it could be argued, are not really cumulated distinct features but values of a complex person/number feature (Haspelmath 2009). We briefly discuss the criteria adduced by Ackerman and Stump. Criterion I: Feature intersection In its simplest form the property of feature intersection (defining our ‘intersective periphrasis’) is the strongest argument for recognizing periphrasis as a distinct linguistic phenomenon. However, we should point out that there are non-standard ways in which features may intersect, and we will touch upon these as we proceed. Criterion II: Non-compositionality The main difficulty with applying a criterion of non-compositionality is that it hinges on a prior definition of morphosyntactic properties and their modes of combination. The underlying intuition is based on unification: if two expressions bear contradictory values for a feature, say, ‘[number:singular]’ and ‘[number:plural]’ then their representations cannot be unified. However, suppose we regard singular number as an unspecified default value: ‘[number:∅]’ or whatever. Then we could argue that the specification ‘[number:plural]’ (or ‘[number:singular]’ for that matter) is compatible with the specification ‘[number:∅]’ and there would be no clash to impede unification. A subtle example of this problem can be seen by comparing our example of the Spoken French past tense periphrasis with the analysis in Co Vet (2007), which argues that the semantics of the tense–aspect system is, after all, compositional. From there the authors argue (implicitly) that the feature system underlying the inflectional paradigms is also compositional in the sense required by Criterion II. Even if we were to grant the semantic aspects of the analysis, it seems to us that their analysis fails to address fully the purely morphological fact that there is an opposition between present and past tense and between simple and perfect aspect that cannot be described by a coherent system of morphological features. Whatever the correct analysis turns out to be, the example illustrates very clearly some of the complexities behind Criterion II. Criterion III: distributed exponence Criterion III, distributed exponence, seems to us to be less successful as a criterion for periphrasis. For instance, in Indo-European languages we expect verbs to inflect for person/number and adjectives for number/gender. In many Indo-European languages

periphrasis and inflection

215

tense, mood, or voice constructions are expressed by a copular auxiliary together with an adjectival, participial verb form, and predictably the verbs inflect for person/number while adjectives inflect for number/gender. This is formally speaking an instance of distributed exponence of the person and gender properties, and hence criterial for a periphrase. If anything, however, it reflects less than full grammaticalization. An illustration of this is found in the South Slav languages Bulgarian and Macedonian. Unusually for Slavic languages, they have both recently developed a ‘have’ perfect construction. The verb ‘have’ combined with an object modified by passive participle has been reinterpreted (as it was in Romance and Germanic), schematically: I have [five books ordered] ⇒ I [have ordered] five books, as in the Bulgarian example (11): (11)

Imam por˘aˇc-an-i pet knigi have.1sg order-passptcp-pl five books ‘I have ordered five books.’ or ‘I have five books ordered.’

(Bulgarian)

The grammaticalization process has gone further in Macedonian than in Bulgarian (where it is a very recent development), so that the Macedonian ‘have’ perfect is now essentially a stylistic variant of the native Slavic resultative perfect with the ‘be’ auxiliary and a resultative (active voice) participle (the l-participle). There are several indications that the Bulgarian construction is not yet a fully grammaticalized periphrastic perfect. First, it is found only with transitive verbs. Second, it is restricted to non-stative predications (cf 12b), preferably durative states (as in 13a): (12)

a.

b.

(13)

a.

Toj e obiˇcal pet ženi he be.3sg love.lptcp.m.sg five women ‘He has loved five women.’ ∗

Toj ima obiˇcani pet ženi he have.3sg love.passptcp.pl five women ‘(intended) He has loved five women.’

(Bulgarian)

Toj ima speˇceleni pet maˇca he have.3sg won.passptcp.pl five matches ‘He has won five matches.’

b. ? Toj ima ritnati pet topki he have.3sg kicked.passptcp.pl five balls ‘He has kicked five balls.’

(Bulgarian)

The Macedonian ‘have’-perfect fails to show its orgins as a predicative adjective construction, in that the participle no longer agrees in number/gender with the object noun phrase, but rather takes an invariable form (syncretic with the neuter singular form), much as in Modern Greek (Migdalski 2006: 136). (14)

a. Ja imam skinato mojata nova košula 3sg.f.acc have.1sg tear.passptcp.n my.def.f.sg new shirt[f] ‘I have torn my new shirt.’

216

andrew spencer and gergana popova b. Go imam skinato mojot nov kaput 3sg.m.acc have.1sg tear.passptcp.n my.def.m.sg new jacket[m] ‘I have torn my new jacket.’ (Macedonian)

Moreover, it is largely unrestricted lexically and semantically (Migdalski 2006: 134–6) and can be used with unaccusative (15a), unergative predicates (15b), or even with reflexive verbs (15c, d): (15)

a. Gostite imaat dojdeno guests.def have.3pl arrive.passptcp.n ‘The guests have arrived.’ b. Goce Delˇcev ima spieno tuka Goce Delˇcev have.3sg sleep.passptcp.n here ‘Goce Delˇcev has slept here.’ ´ c. Veke se imam izmieno already refl I.have wash.passptcp.n ‘I have already washed myself.’ d. Brodot se ima udreno vo karpite ship.def refl have.3sg hit.passptcp.n in rocks.def ‘The ship hit the rocks.’

(Macedonian)

What we see in Macedonian is thus the eradication of distributed exponence, and this is a symptom of greater grammaticalization and hence of a periphrase. Spencer (2008, 2012b) argues that the real reason why the Udmurt construction provides evidence of periphrasis is because the distribution of properties goes against the normal morphosyntax of the language. He points out that exactly the opposite situation occurs in Japanese conjugation. In that language there is no agreement, so that the morphosyntactic (functional) properties of a clause by default show distributed exponence. Japanese verbs inflect for tense (present/past), status (plain/polite), and polarity (affirmative/negative). Any two marked values of these properties can be expressed morphologically (i.e. {past, polite}, {past, negative}, {polite, negative}). However, the combination {past, polite, negative} cannot be expressed by a single word but has to be realized by combining the (default, non-past), polite, negative form of the lexical verb with the past, polite, affirmative (default non-negative) form of the copula/auxiliary desu: tabemasen desita ‘didn’t eat (polite)’. A clear signal that this is a periphrase is the fact that the polite property is realized on both components: everywhere else in the grammar politeness can be realized only once, on the head verb of the construction. Thus, what makes the Japanese construction periphrastic is multiple exponence, that is, the failure of the expected (default) distributed exponence. Indeed, Spencer (2012b) argues that multiple exponence might be a better general criterion for periphrasis cross-linguistically than distributed exponence. However, perhaps the real criterion should be something like ‘if the morphosyntax of a construction contravenes the normal patterning for that language then the construction is (likely to be)

periphrasis and inflection

217

a periphrase’. This would capture the intuition that periphrastic syntax is more highly grammaticalized than normal syntax. Spencer (2006a: 291) (see also Spencer 2003) mentions two further possible criteria for periphrasis. The first is ‘underexhaustivity’, when a function word in a periphrastic construction has fewer inflected forms than would be expected (for instance, it might have fewer inflected forms than a homophonous or polysemous word in a nonperiphrastic construction). The intuition here is that arbitrary restrictions on feature realization are typical of morphological paradigms and not of syntactic constructions. The second criterion is ‘overexhaustivity’. This refers to a much less common phenomenon, in which a periphrastic construction gives rise to more forms than would be expected from the normal rules of syntax. Spencer cites the case of the evidential (Narrated form) periphrasis in Bulgarian in which the auxiliary plus participle combination expresses an evidential (Narrated) mood. This construction can re-apply to the auxiliary part of the construction to give a doubly evidential form, the Renarrated mood (with somewhat complex semantics), rather as though in English we could re-apply the perfect construction to the perfect auxiliary to get has had seen (whatever that would mean). In the final sections we develop the question of criteria for periphrasis, looking at a variety of problems that are encountered when we try to distinguish between periphrases and morphological constructions, clitic constructions, syntactic constructions, and intricate combinations of all three. We then conclude with a brief discussion of criteria for periphrasis from a typological standpoint.

9.7 Periphrasis or Morphology

.............................................................................................................................................................................

As mentioned before, feature intersection is the only really reliable diagnostic for periphrasis. However, there is an interesting problem in defining what we mean by ‘feature intersection’ when we consider the philosophy underlying that criterion. The idea is that we define a construction as periphrastic if we would have expected morphological, synthetic expression of that grammatical property given the rest of the grammar. Where it is easy to distinguish word forms from each other, and where word forms bear clearly identifiable inflections, it is relatively easy to determine the structure of the inflectional paradigms and determine whether or not a cell is filled by a synthetic form. However, we often find that function words can become morphologized in certain of their contexts of use, while remaining distinct word forms in other contexts. Such function words generally first become clitics and then undergo further grammaticalization to become more like affixes. But if this process of affixation is not found with all uses of that function word then we have an alternation between morphological expressions and genuine multi-word expressions, and the latter assume the character of periphrases.

218

andrew spencer and gergana popova

9.7.1 Fused (Portmanteau) Auxiliary Constructions As is well-known, English unstressed auxiliaries are enclitics. However, when they are attached to subject pronouns functioning as the sole subject of the clause, they fuse with the pronoun and trigger idiosyncratic allomorphy, forming, in effect, a portmanteau: you are (/juː ɒː/) ⇒ you’re (/jɔː/)13 If we regard such fused auxiliaries as affixes then we should treat forms such as you’re as pronouns inflected for tense (and agreeing with their own stem). But in that case, the analytic auxiliaries should be viewed as periphrastic constructions, although they will then be periphrastic forms of the pronominal lexeme, not periphrastic forms of the verb lexeme. This suggests that we need to distinguish affixation (and hence inflection) proper from portmanteaux which develop from frequent collocations, such as you’re. We will discuss two distinct reflexes of portmanteau morphology. First, there are cases of on-going language change in which a construction built out of one of more function words becomes morphologized but only in certain registers or only in certain lexical or grammatical conditions. A simple, familiar example of this is the phenomenon of ‘fused prepositions’ in German and Romance languages. In general, a preposition plus definite article construction is expressed by two distinct words, which can, in principle, be separated (say, by pauses). However, certain combinations obligatorily or optionally fuse to give a single word form. Obligatory fusion is found in French, for instance, with the prepositions de ‘of, from’ and à ‘to’ followed by the m.sg or pl definite article: de+le = du, de+les = des, à+le = au, à + les = aux. With the f.sg article we get no morphophonological fusion: de+la, à+la. Other prepositions show no fusion: pour le ‘for the.m.sg’, sous les ‘under the.pl’. Examples such as these raise interesting questions about the morphophonological status of the prepositions themselves (are they clitics or phrasal affixes, for instance? See Miller 1992 for discussion). The point of the examples is that we have single words forms which express the lexical/functional meaning of the preposition together with the functional features of gender and number. They therefore count as an unusual instance of inflection (in which a word receives inflections which properly belong to a different syntactic head). However, they are restricted to a very small number of forms (in French). The question therefore arises as to whether we should treat all other preposition plus article combinations as periphrastic, on the grounds that four combinations (out of the six possible with a/de) are (in some sense) inflectional. While we might want to dismiss the French examples as a species of ‘anti-periphrasis’ (see Section 9.3), the situation in other Romance languages is more complex. In Portuguese and Italian we see a greater number of forms with a greater number of prepositions. Italian has six such prepositions, though some of the forms are oldfashioned (Maiden and Robustelli 2000: 63). Literary Welsh, discussed in Section 9.3, 13

Standard British English pronunciation.

periphrasis and inflection

219

has seventeen inflecting prepositions. At which point do we say that the default category is that of inflecting preposition (at least for monosyllabic prepositions) and that non-inflecting monosyllabic preposition plus article combinations are periphrastic? A particularly striking example of such morphologization is seen in the Serbian/ Croatian future marker. When stressed, this is a full form auxiliary verb htˇeti, homophonous with the lexical verb ‘to want’. This auxiliary can combine with the infinitive form of the verb (reflecting its historical origins as the complement of the verb ‘want’): (16)

Ho´ce ˇcitati knjigu fut.3 read.inf book.acc ‘He/she/they will read the book.’

However, when unstressed, the auxiliary takes the form of a Wackernagel clitic, ´cu, ´ceš etc. (17)

ˇ knjigu c´ e Citati read.inf fut.3 book.acc ‘He/she/they will read the book.’

The future clitic often appears immediately after a clause-initial infinitive form lexical verb. In Western dialects (e.g. Croatian) especially, it is then often fused with the infinitive form to create what is effectively a future suffix: (18)

ˇ ce Cita´ knjigu read.inf.fut.3 book.acc ‘He/she/they will read the book.’

Thus, one and the same formative can effectively appear either as an affix, a Wackernagel clitic, or a full form auxiliary. One plausible way of extending the criterion of feature intersection is to say that where a multi-word construction actually alternates with a morphological, affixal variant, then this is good evidence that the multi-word construction itself is a periphrasis.

9.7.2 Hausa ‘Genitives’ Hausa provides the second type of alternation between morphological and periphrastic expression. The standard way to express possession in Hausa is by means of a possessive particle, or ‘linker’, which signals the number/gender of the possessor. This linker has the form ta if the head noun (possessed noun) is feminine singular and na otherwise (i.e. when the possessed noun is masculine or plural). These examples are taken from Newman (2000: 300): (19)

a. [bák¯a]M na mahá˜rb¯ı bow[m] of.m hunter ‘the hunter’s bow’

220

andrew spencer and gergana popova b. [kibiy `ā]F ta mahá˜rb¯ı bow[f] of.f hunter ‘the hunter’s arrow’

The linkers na/ta are in construction with the possessor noun, mahá˜rb¯ı, not with the possessed nouns. For instance, it is possible to use the bare forms na mahá˜rb¯ı, ta mahá˜rb¯ı to mean ‘the hunter’s one’, without the possessed noun. However, the linker frequently attaches as an enclitic to the previous word, that is, to the possessed noun head with which it agrees in gender/number. When this happens na surfaces as =n and ta as =˜r.14 (20)

a. báka=n mahá˜rb¯ı bow=of.m hunter ‘the hunter’s bow’ b. kibiyà=˜r mahá˜rb¯ı bow.of.f hunter ‘the hunter’s arrow’

There are good reasons to believe that the default form of the construction is the multiword variant in which the linker is an autonomous function word, and not the variant with the fused clitic/affix. For instance, the fused form is not found when the head noun is otherwise modified, as in (21a) (compare this with 21b) (Newman 2000: 301): (21)

a. r`īg¯a baƙ¯a ta Lawàn gown[f] black of.f Lawan ‘Lawan’s black gown’ b. r`īga=˜r Lawàn baƙ¯a gown[f]=of.f Lawan black ‘Garba’s two bicycles’

The structure of the basic Hausa possessive construction is therefore (22a) alternating with a morphologically fused form (22b): (22)

a. Possessed b. [Possessed=linkeri ]

[linkeri Possessori ] Possessori

(The Hausa construction is reminiscent of the Danish definite marker in this respect.) In Hausa, then, we see a possessor construction which has the hallmarks of a multiword construction, involving an independent linking element in construction with the possessor noun, but agreeing with the possessed noun. Under certain circumstances, however, this construction has been morphologized so that we can plausibly speak of a construction in which the possessed noun agrees with the possessor noun (although the agreement marker may be best thought of as a phrasal affix rather than 14 We have constructed examples (20) on the basis of the examples given by Newman (2000: 302), bàkan gizò ‘rainbow (=‘the bow of Gizo-the-trickster’)’ and kibiyà`r y¯ar ō` ‘the boy’s arrow’.

periphrasis and inflection

221

a head-marking affix). Thus, we have a kind of feature intersection, though not the one seen in examples like Latin conjugation. Again, we can regard such alternations as supporting the analysis of the multi-word alternant as a periphrase.

9.8 Periphrasis or Clitics

.............................................................................................................................................................................

The existence of clitic systems raises difficult questions for the characterization of periphrasis, because clitics have some of the properties of affixes and some of the properties of function words. On the one hand, if we regard a given clitic-type formative as essentially a word then we are saying that the constructions it enters into are multi-word constructions and hence candidates for periphrases. On the other hand, if we regard a clitic-like element as essentially an affix then we increase the space of constructions that we regard as being realized morphologically. This, consequently, increases the possibility of feature intersection involving related, but clearly syntactic, constructions. Thus, the existence of clitics raises the probability of finding periphrastic expression, but in different directions, so to speak. These issues have not been seriously dealt with in the literature on periphrasis so our remarks will be provisional. ˘ ‘be’, as well as direct/indirect object In Bulgarian, the present tense of the copula sam pronouns, form a clitic cluster, placed after the first phrase in the clause. The copula is used to form the perfect tense series with the l-participle form of the verb (here dala ‘gave’): (23)

a. Dala s˘am mu gi give.ptcp.f be.prs.1sg cl.3sg cl.3pl ‘(I) have given them to him.’ b. Az s˘am mu gi dala I be.prs.1sg cl.3sg cl.3pl give.ptcp.f ‘I have given them to him.’

(Bulgarian)

The cluster cannot appear clause initially: (23)

c.

∗

S˘am mu gi dala be.prs.1sg cl.3sg cl.3pl give.ptcp.f

The negation element ne and the invariable future tense particle šte also enter the clitic cluster, appearing at its leftmost edge in the order ne šte. The linear order is represented schematically in (24): (24)

Bulgarian clitic cluster: Neg < Fut < Aux < Dat < Acc < 3sgPresAux

222

andrew spencer and gergana popova

Clitics in the cluster appear pre-verbally if possible, and adjacent to the verb. They appear to the right of the verb if there is no available host to the left and if none of the clitics in the cluster can appear clause-initially.15 (25)

a. Az ne s˘am mu gi dala I cl.neg be.prs.1sg cl.3sg cl.3pl give.ptcp.f ‘I haven’t given them to him.’ mu gi dala b. Az šte s˘am I fut be.prs.1sg cl.3sg cl.3pl give.ptcp.f ‘I shall have given them to him.’ c. Az ne šte s˘am mu gi dala I cl.neg fut be.prs.1sg cl.3sg cl.3pl give.ptcp.f ‘I shall not have given them to him.’

(Bulgarian)

From these data we can conclude that the clitic cluster occupies (i) second position in the clause (is a Wackernagel cluster), (ii) is proclitic to a verb form. Together these properties guarantee that the cluster either appears immediately after a clause-initial verb form, or appears sandwiched between the first phrase and the verb. The principle that the cluster should always be adjacent to a verb is familiar from the history of Romance clitics and is often referred to as the Tobler–Mussafia law (Halpern 1995; Anderson 2005; Spencer and Luís 2012). The picture is complicated, however, by the fact that both ne and šte are able to appear in absolute clause initial position and hence to host the rest of the clitic cluster: (26)

a. Ne s˘am mu gi dala cl.neg be.prs.1sg cl.3sg cl.3pl give.ptcp.f ‘I haven’t given them to him.’ b. Šte s˘am mu gi dala fut be.prs.1sg cl.3sg cl.3pl give.ptcp.f ‘I shall have given them to him.’

(Bulgarian)

In this respect ne, and šte apparently behave like autonomous words and not like clitics. Similar problems may arise with any system in which function words have been partially grammaticalized as clitics (and in some cases where fully fledged function words alternate with affixes, as in Hausa genitives, or English subject pronouns with reduced auxiliary suffixes). The aberrant linearization seen in (24) is typical of inflectional, affixal systems and is unheard of in genuinely syntactic constructions (see Spencer and Luís 2012: ch. 8 for discussion). This fact has motivated treatments in which the cluster is treated as an essentially morphological construct (Legendre 2001; Spencer 2001). If we pursue 15 It is not usual for the future to be negated by ne, as in (25c), however. The periphrastic construction with the invariable particle njama followed by a finite clause (Section 9.9) is the preferred way of forming the negative future.

periphrasis and inflection

223

such reasoning we have to conclude that the perfect tense forms in (23) and also the future tense forms are synthetic and not periphrastic, much less purely syntactic constructions. It is undoubtedly controversial to treat the Bulgarian clitic cluster as a kind of morphological phenomenon (although it is hardly any less controversial to treat it as a kind of syntactic phenomenon either). Moreover, there are models of morphosyntax which permit a dual categorization of partially grammaticalized elements of this sort, under which they are affixes at one level of description and syntactic words at another level (see in particular Sadock’s 1991 model of Autolexical Syntax for an analysis of clitic systems which explicitly adopts such a dual categorization approach). Therefore, it remains open to the analyst to argue that the clitics count as ‘words’ for syntactic purposes, and hence that they participate in multi-word constructions. In that case, depending on the specifics of the constructions, we might want to say that the construction is a periphrasis, albeit one which is realized by weak or deficient syntactic terminals, which can even take the form of affixes. However, the conceptual problem is thrown into further relief when we consider the almost identical clitic system of Macedonian, a language which is mutually intelligible with Bulgarian. Macedonian has essentially the same clitic cluster as Bulgarian (except that both 3sg and 3pl forms of the present tense of sum ‘to be’ appear at the ´ (cognate with Bulgarian right edge). In particular, it has an invariable future clitic, ke šte). In contrast to the Bulgarian cluster, the placement of the Macedonian cluster is defined exclusively with respect to the verb (whether the lexical verb or some non-clitic form of sum). If the verb is finite, the cluster is a proclitic and if the verb is non-finite (a gerund or imperative form) the cluster is enclitic. As a result it is perfectly possible for any form of the cluster to be in absolute clause initial position. This is essentially the pattern found with modern Romance languages such as French, Spanish, and Italian. A number of linguists have explicitly argued that these modern Romance clitic systems are actually types of (loose) affixation and not true clitic systems (see Bonami and Boyé 2007 for French, Monachesi 1999 for Italian, and Spencer and Luís: 2012: 89–90 for general discussion. Halpern 1995: 26–32, 183–5 explicitly draws a comparison between Bulgarian and Medieval Romance on the one hand and Macedonian and Modern Romance on the other). In fact, there are even better grounds for treating the Macedonian clitic cluster as a kind of loosely cohering affix string. This is because the pronominal elements of the string are obligatory cross-referencing elements for all indirect objects and for definite direct objects. (In Bulgarian, it is only the indirect object pronominals that are obligatory, and doubling of a direct object by the clitic is subject to rather complex discourse-level factors). As a number of authors have pointed out (for example, Lyons 1991; Spencer 1991), the Macedonian pronominal clitics are effectively agreement affixes. But, since they are part of a strict cluster which includes the negation and future tense particles, we have to conclude that negation and future tense are also expressed morphologically in Macedonian. The upshot of the Macedonian patterning is that we have rather more reason to adopt a morphological analysis of

224

andrew spencer and gergana popova

the clitic cluster than we have for Bulgarian. This is important because the Macedonian future tense exhibits the same periphrastic alternation in the negative as the Bulgarian future tense, as discussed in the next section. It is difficult to draw more general conclusions about the relationship between clitic constructions and periphrases because of the dual nature of clitics. If we chose to focus more on the word-like properties of clitics then clitic constructions, being highly grammaticalized, should surely be regarded as a subtype of periphrasis (although one which is often very idiosyncratic in its syntax). If, however, we chose to focus on the affixlike properties of clitics, then we are simply dealing with a (partially) morphologized construction and it does not make much sense to think of this as a periphrase. The real question will be whether the construction retains enough syntax-like properties to warrant treating its formatives as word-like, that is, as independent syntactic terminals. In any event the properties of clitic constructions only serve to emphasize the need to distinguish between ‘ordinary’ syntax and the often highly idiosyncratic syntax exhibited by periphrases.

9.9 Periphrasis or Syntax

.............................................................................................................................................................................

One way in which we might justify treating a multi-word construction as inflectional periphrasis is when that construction is used as a variant of a synthetic word form, either with or without additional stylistic nuances. This is reported in Ancient Greek for the perfect active tense series, for instance, which can be expressed either by an inflected form or by a combination of ‘be’ with a perfect active participle (Weir Smyth 1920: 182). Another case in which we find synonymy between synthetic and analytic constructions is the Catalan past tense or preterite. The synthetic preterite for a regular first conjugation verb such as cantar ‘sing’ is cantí, cantares, cantà, cantàrem, cantàreu, cantaren (Gili 1967: 57; Wheeler et al. 1999: 301). However, the same meaning can be expressed periphrastically, by combining the infinitive of the lexical verb with the present indicative forms of the verb anar ‘go’, used as an auxiliary: vaig cantar ‘I sang’, vas cantar ‘you(sg) sang’ and so on (Gili 1967: 65; Wheeler et al. 1999). Intriguingly, the auxiliary is not actually homophonous to the lexical verb anar. In the 1pl, 2pl the auxiliary forms are vam, vau while the lexical verb has the forms anem, aneu. Wheeler et al. (1999: 343) explicitly state that the synthetic and periphrastic forms are synonymous and difficult to distinguish even on the basis of style or register. The point about such examples is that the multi-word expressions can only express a very specific combination of inflectional properties, so we are not dealing with parallel lexemes which just happen to have similar meanings for some of their forms.16 16

Such examples are of wider interest for the theory of inflection because they represent inflectional doublets or ‘overabundant expression’ (Thornton 2012), something which is supposed to be ruled out

periphrasis and inflection

225

Bulgarian provides a number of interesting further examples of multi-word inflectional constructions. Here we briefly summarize three aspects of these candidates for periphrasistic status: (i) alternations between clitic and non-clitic auxiliaries, (ii) subordinate clauses as parts of periphrases, and (iii) nested periphrases. ˘ ‘be’ in BulgarWe said in the previous section that the auxiliary/copular verb sam ian is a clitic. However, this is only true of its present tense forms. In all other forms it behaves like an autonomous non-clitic syntactic word in its own right. Thus, the past perfect employs the imperfect tense form of the copula, which is not a clitic, and therefore can (and often must) appear clause-initially: (27)

a. Bjax dala statiite na studenta. be.impf.1sg give.ptcp.f be.prs.1sg papers.def to student.def ‘(I) had given the articles to the student.’ (Bulgarian)

If we regard the clitic construction as morphological then we must treat the nonclitic auxiliaries as part of an intersective periphrase. The clitics define a morphological paradigm space so that the property of ‘perfect’ is now a morphological (properly inflectional) property. Therefore, the past perfect construction is the intersection of two morphologically expressed properties, perfect and past. On the other hand, if we regard the clitics as a kind of reduced word, then there will be less motivation for treating the non-clitic auxiliaries as part of a periphrase. The perfect property would no longer be a morphological (properly inflectional) property and so the feature combination past perfect would no longer define an intersection of features. We leave this question unresolved since it involves a host of poorly understood questions.17 However, the question serves well to illustrate the kinds of conceptual problems we face when characterizing periphrasis. For Macedonian, on the other hand, it seems much more reasonable to recognize the clitic system as a type of morphology. In that case its perfect, future, and negative constructions are not periphrastic but synthetic. We now turn to point (ii), the existence of periphrases expressed by subordinate clauses. Bulgarian is unusual in that it includes fully fledged finite subordinate clauses in its system of obligatory grammatical contrasts and these are arguably periphrases (Spencer 2003; see also Spencer 2012b for other examples of this phenomenon in the context of periphrastic negation). The negation of the future, in particular, is expressed by means of an invariable particle njama (derived historically as the fused negative form of the verb ima ‘have’), which combines with a finite subordinate clause containing the content lexeme. In (28a) we see an affirmative polarity clause, in which future tense is expressed by the clitic šte and s˘am mislila is the present perfect tense form of the lexeme mislja ‘think’. In (28b) we see the same sentence in the negative polarity, on certain paradigm-based models of inflection Stump (2001b) which treat inflectional rules as strictly mathematical functions and not relations. 17 In Popova and Spencer (2012) we treat the clitics as syntactically represented words, in part for ˘ inflects in the manner of fully fledged expositional convenience, in part because the auxiliary clitic sam verbs.

226

andrew spencer and gergana popova

in which da s˘am mislila po v˘aprosa is a subordinate clause in the present perfect tense, introduced by the marker/subordinating conjunction da: (28)

a. Az šte s˘am mislila po v˘aprosa. I fut be.1sg.prs think.ptcp.f on matter.def ‘I will have thought about the matter.’ b. Az njama da s˘am mislila po v˘aprosa. I have.not da be.1sg.prs think.ptcp.f on matter.def ‘I will not have thought about the matter.’

(Bulgarian)

Recall that we remained non-committal over the status of the future clitic. Suppose we adopt the radical morphology analysis under which all of the members of the clitic cluster are actually morphological exponents. In that case, the negated future with njama provides a clear instance of an intersective periphrasis. Suppose, however, that we reject the morphological analysis and treat the future particle šte as a syntactically represented word. In that case we can note that the njama construction is featurally non-compositional, in the sense that the property [Tense:future] cannot be projected from any of the components of the construction (and the property [Polarity:negative] is embedded inside a complex word form). By Criterion II of Ackerman and Stump (2004) we would therefore take the negated future to be a typical instance of a periphrasis. That conclusion would hold independently of our analysis of the affirmative future form. Again, the comparison with Macedonian is interesting, for that language exhibits exactly the same kind of patterning for its negative future construction. Kramer (1999: 98) provides the following examples illustrating the ‘compositional’ negation of the ´ as well as the more usual periphrastic future by prefixing ne to the future particle ke, pattern with the negative impersonal verb nema: (29)

´ im a. Ne ke ja ˇcitam knigata na decata neg fut 3pl.dat 3sg.f.acc read.1sg book.def to children.def b. Nema da im ja ˇcitam knigata na neg.impersonal da 3pl.dat 3sg.f.acc read.1sg book.def to decata children.def ‘I will not read the book to the children.’ (Macedonian)

Mišeska Tomi´c (2006: 449) reports that the future negated with ne is restricted semantically compared with the nema construction, in that the future with ne has to involve what she calls a ‘presuppositional’ component, that is, an element of supposition or evidentiality. Only the nema construction is available for straightforward negative assertions. We concluded earlier that the clitic cluster in Macedonian can only sensibly be analysed as a kind of loosely cohering affix string. This means that the affirmative future

periphrasis and inflection

227

form is synthetic, and hence that the property [Tense:future] defines part of a morphological paradigm. However, the negation of the future with nema is a multi-word construction on any analysis, just as in Bulgarian. Therefore, we must conclude that the negative future in Macedonian is an instance of featurally intersective periphrasis. Finally, we come to an obvious point about periphrases, raised in Popova and Spencer (2012), which has been largely neglected (it is not mentioned for instance in Ackerman and Webelhuth’s (1998) otherwise very thorough discussion of the theoretical implications of verbal periphrases). We may well have more than one periphrastic construction in a language and, just as purely morphological properties can be intersected with each other, we expect the properties expressed by periphrases to be intersected. For lexeme-based realizational models that wish to incorporate a notion of periphrasis as an exponent of inflectional properties, this raises the implementational question of how exactly we define such nested periphrases. In fact, we have, arguably, already seen simple examples of nested periphrases in Bulgarian: if we take the šte future to be periphrastic and if the perfect is periphrastic then the combination of the two properties, the future perfect, will a fortiori be a periphrase. Thus, example (28a) above is the combination of the perfect periphrase with the future periphrase, while example (28b) is a combination of three periphrases: negation(future(perfect)).

9.10 Periphrasis and Typology

.............................................................................................................................................................................

The problem of establishing a typology of periphrasis is hampered by the dearth of cross-linguistic comparisons of the phenomenon and that dearth is largely the result of a lack of clear, agreed, criteria for what counts as periphrastic. The principal comparative discussion tends to be found in handbook articles such as Haspelmath (2000) or Spencer (2006a), otherwise comparison tends to be limited to Latin and Greek or, say, Romance languages and classical languages (mainly Latin and Greek). We briefly mention here three contrasting sets of typological claims from the recent literature. Brown et al. (2012) offer a somewhat different perspective on periphrasis within the context of Canonical Typology. In Canonical Typology (Corbett 2006, 2007; see also the papers of Chumakina and Corbett 2012, Brown et al. 2013) we set up a series of (hopefully uncontroversial) canonical criteria for membership of a category or construction type. The canonical criteria do not have to be frequently instantiated in actual languages and the criteria themselves do not define any kind of ‘prototypical’ category or construction, but rather define a logical space in which actual linguistic expression types deviate from the canonical in strictly definable ways. Brown et al. argue that we can define an (abstract) notion of ‘canonical periphrasis’, which is simultaneously a canonical instance of (functional) syntax while also being a canonical exponent of paradigmatic inflection. Being an exponent of inflection, a canonical periphrase has to realize uncontroversially inflectional morphosyntactic

228

andrew spencer and gergana popova

features and also has to observe the criterion of feature intersection. Being canonical functional syntax, the periphrase has to have the same, regular, syntax as other syntactic constructions expressing functional information or relations. This criterion does not figure in Ackerman and Stump’s (2004) list or in accounts based on prototypes because periphrases are typically non-canonical with respect to syntax. However, a canonical periphrase lacks such idiosyncrasy (including non-compositionality and distributed exponence, where this is non-canonical for the language) and for that reason those periphrases which shade into clitic constructions are non-canonical. The Latin passive perfect periphrasis is close to canonical on this analysis (although not entirely so, given that it is not entirely compositional). Bentein (2011) discusses the various instantiations of the Ancient Greek perfect periphrasis briefly introduced in Section 9.3.2. He traces the gradual grammaticalization of various types of periphrastic perfect in order to determine when we can reliably describe the construction as ‘periphrastic’, in rather the way that we have talked about the Bulgarian vs. Macedonian have-perfects. He argues for three sets of prototypical properties for periphrasis. The first two correspond to the universally accepted criteria under which the periphrasis must express grammaticalized (inflectional) properties (‘semantic criterion’) and must respect feature intersection (‘paradigmatic criterion’). He also argues that the periphrasis has to have more rigid syntax, specifically the auxiliary plus participle combination should be more tightly cohering than in normal syntactic constructions (which show considerable word order freedom in Ancient Greek). This is a prototypical criterion but for Brown et al. (2012) it is not a canonical one. A more systematic and formalized exploration of such ideas is put forward by Bonami and Samvelian (2009, 2014). They discuss several multi-word constructions in Persian: (i) passive, (ii) perfect, (iii) future, and (iv) progressive. The passive is argued to be a purely syntactic construction (and hence not periphrastic), indeed, a kind of productive light verb construction with the auxiliary form of the verb šodan ‘become’. By contrast the perfect forms enter into a paradigmatic relation with a synthetic ‘complex present perfect’ construction and thus constitute a case of feature intersection. The future and progressive represent different degrees of grammaticalization and are periphrastic to a greater or lesser degree. The present perfect is formed from the perfect participle and a bound form of the copula/auxiliary which has become a suffix. Thus, from the verb foruxtan ‘sell’, past stem foruxt-, past participle foruxte we obtain foruxte=am ‘I have sold’. In the past, however, the copular/auxiliary is an inflected form of the stem bud-, thus foruxte budam ‘I had sold’. Although the participle and past form of the auxiliary cannot for the most part be separated, it is possible for the participle to be fronted, as in Bonami and Samvelian’s (2014) example (45), given here as (30) (see also Bonami and Samvelian 2009, example (23)): (30) Foruxte fekr nemikonam bâˆsad in tâblo-râ sold thought neg.light.verb be.sbjv-3sg this painting-do ‘I don’t think that s/he has sold this painting.’

periphrasis and inflection

229

Such fronting is entirely excluded in the present perfect. Thus, Bonami and Samvelian argue that the auxiliary and the participle are separate words so that the past perfect is a multi-word expression. Since it is in paradigmatic contrast with the present perfect, and since the present perfect is a single inflected word form, we have an instance of feature intersection, thus, the past perfect is periphrastic. The periphrasis is ingeniously coded in the HPSG analysis by treating the periphrastic forms as realizations of the lexical verb lexeme, V, (as defined by the lexical index or LID) but realizing the construction as a property of the form of the auxiliary selecting the V’s participial form as a valent. This is shown schematically in (31) for the construction foruxte bud ‘had sold’: (31)

Persian: foruxte bud ‘(s/he) had sold’ V′

 1

Vpastpart  LID PHON



 sell foruxte



Aux3sg past  LID   PHON ARG-ST

sell bud

  

The auxiliary verb is not an autonomous lexeme in this construction, any more than an inflectional suffix is an autonomous lexeme. This is achieved in the morphology as part of the paradigm function. Bonami and Samvelian apply the model of inflectional morphology developed by Stump (2001b), Paradigm Function Morphology. The machinery of Paradigm Function Morphology is extended slightly by allowing morphological rules to make reference to argument structure as well as phonological representations. A rule of referral serves to realize the perfect aspect property by referring us to the phonological representation and the argument structure of the auxiliary verb. That auxiliary verb has a lexical specification requiring it to take the past participle of the original verb lexeme as one of its arguments. Thus, the referral rule indirectly introduces the perfect participle of the verb lexeme as well as the perfect auxiliary. Bonami and Samvelian’s work is additionally interesting because they argue that the future and the progressive represent constructions which are not canonically syntactic (as is the passive) but which are not canonically periphrastic either, in that they do not involve intersection with morphological paradigms. Their analysis neatly teases out the relevant distinctions in a formally explicit way and offers potentially very interesting insights into the types of periphrastic construction found cross-linguistically.18 18

The verb system of the Tajik variety of Persian is even more interesting, in that it has innovated another periphrastic progressive based on the perfect participle istoda of the verb istan ‘stand’ and has extended the periphrastic evidential to include the whole tense/aspect series. See Rastorgueva and Kerimova (1964: 71–97) for detailed discussion of the evidential and Windfuhr and Perry (2009: 446–69) for an explicit comparison with Iranian literary standard Persian.

230

andrew spencer and gergana popova

9.11 Conclusions

.............................................................................................................................................................................

The traditional notion of ‘periphrastic construction’ has been given new life within the inferential-realizational class of models of inflectional morphology, which are based on the notion of the structured paradigm. In its purest (canonical) form a periphrastic construction is a multi-word construction which is clearly syntactic in form (and fully conforms to the language’s syntactic principles) but which serves the function of realizing sets of properties that are otherwise realized morphologically , that is, ‘intersective periphrasis’ (Brown et al. 2012). Inflectional periphrases thus reflect a paradigm structure which is deviant with respect to morphology. If we were to limit the notion ‘periphrasis’ to canonical intersective cases, however, then many constructions traditionally thought of as periphrastic would not fall under our definition, including the English progressive/perfect and passive periphrases. Various authors have therefore proposed that there are several other diagnostic (or prototypical) properties of periphrases (it is not clear whether we should regard them as sufficient conditions). One of these is idiomatic or non-compositional interpretation (Criterion II of Ackerman and Stump, 2004). Another is idiosyncratic syntax compared with the rest of the (functional) syntax of the language. This may involve what Ackerman and Stump call distributed exponence (cf. their Criterion III). However, we have argued that what is really at stake is a deviation from the language’s syntactic norms. In some cases, this may mean that other properties such as multiple exponence are diagnostic of periphrasis (the Japanese polite past negative) and in other cases it may be the loss of distributed exponence that is a diagnostic for periphrasis (Macedonian ‘have’perfects). In addition to the relatively clear situation of feature intersection, periphrasis can interact with morphological patterning in more complex ways, as when a multiword construction such as the Catalan periphrastic past tense with anar alternates with a (near-)synonymous or functionally equivalent morphologized construction. We take this as a further diagnostic for periphrasis. However, it is in the nature of things that periphrastic constructions lie in the territory between ordinary syntax and ordinary morphology and represent different stages of grammaticalization. There is thus no completely definitive characterization of periphrasis which will permit us to automatically identify periphrastic constructions in any given language. We have summarized some of the complex interactions that we find between periphrases and fully morphologized constructions, periphrases and clitic constructions, and periphrases and non-periphrastic syntactic constructions which we think grammarians should take into account when describing a construction as an inflectional periphrasis. In models of grammar that draw a careful distinction between morphological phenomena and syntactic phenomena it is important to be able to recognize these interactions and we concluded with a brief summary of some of the typological implications of taking seriously the notion of periphrasis as a grammatical construction type.

p a r t iii ........................................................................................................

CHANGE ........................................................................................................

chapter 10 ........................................................................................................

DIACHRONY ........................................................................................................

claire bowern

10.1 Situating the Study of Inf lection in Language Change ............................................................................................................................................................................. 10.1.1 Introduction Work on both the origins of morphological inflection and the operations which apply to morphology in language change dates almost to the beginning of the modern era of historical linguistics. Even in the early nineteenth century, for example, Bopp (1816) sought to compare morphology between related languages and to use those comparisons as evidence for linguistic relationships; he also paid considerable attention to arguing for the origin of such inflection in independent words. More recently, a strand of work in morphology has been concerned with the contributions of historical linguistics to the synchronic analysis of morphology. Some things can only be perceived when they move; likewise, some aspects of morphology are most clearly observed when taking into account the evidence of change. Such changes give us insights into how speakers treat the internal organization of word forms within their language, for example. Chafe (2008: 261) makes perhaps the strongest statement of such a view, when he writes ‘I believe it is impossible to understand language design without taking account of language change, seeing historical processes as shaping the fundamental nature of language.’ Those historical processes are two-fold. First, there is change through language acquisition: in the process of acquiring language, children may come to different generalizations about the structure of their language from the generalizations their parents made (see e.g. Andersen 1973; Lightfoot 1991; Aitchison 2003). Secondly, there is a social component to change: for a change to be actualized within a language, a certain number of speakers must converge on the same analysis, otherwise the new change cannot spread through a community (Harris and Campbell 1995; Andersen 2001).

234

claire bowern

The changes that we see in language are also shaped in various ways—by constraints on cognitive processing (Yu 2010), by prior structures in the language (Blevins 2004), and by language production (Ohala 2001). Finally, there is a stochastic element in language change; paths of change are not determinative. All these elements are necessary for a general theory of morphological change. To address all such areas with respect to morphology in a single article, however, would be impossible; moreover, most recent work on morphological change has not situated morphology within a research paradigm that views language change in this way. That is, research which builds on the interfaces between the social, the psychological, and the systemic is rather less common than research concerned with questions such as what types of change are possible or natural. Research has thus focused on questions internal to the linguistic system rather than research which relates language change to speaker behaviour. Here I am concerned in particular with situating current work in morphological change in the wider literature on morphology. I first consider the changes which target morphology, and inflection in particular. That is, are there changes which affect only inflection? Are there areas of morphology which are immune from other types of change, such as regular sound change? Can sound change be conditioned by morphology? These questions are addressed in Section 10.2. Authors have also been concerned with characterizing the nature of morphological change in terms of how ‘natural’ such changes are, what might condition ‘unnatural’ changes, and how this relates to the question of autonomy in morphological change. These are also questions addressed in this section. In Section 10.3, I discuss sources of morphology; how free words are grammaticalized and organized into paradigms; how changes can remodel existing morphology, and how new paradigms can be created. Morphology has played an important role in the proposal and justification of genetic relationships. Some families, such as Algic (Sapir 1913; Goddard 1975), were initially established solely on the resemblances of morphology; in other cases, such as the recent proposal of a Yenisei-Na-Dené family including languages across both sides of the Bering Strait, the establishment of the family relies crucially on the evidence of morphology. In Section 10.4 I turn to an examination of how morphology has been used to establish language families, and the principles and assumptions that underlie this use of evidence (in contrast to sound change or lexical items). There are many diachronic issues in inflection, and there is not space to consider all of them. Some morphological issues have assumed particular prominence in work on individual language families. In Pama-Nyungan, for example, morphology, and in particular nominal inflectional morphology and pronouns, has comprised an important part of the evidence for the establishment of subgroups. Another area that cannot receive attention here (except in passing) is the relationship between morphology and other aspects of language. It is, of course, artificial to consider language change without reference to the social dimensions of change, such as variation within a speech community (see further Kossmann, this volume) and the actualization of change through a group of speakers (Timberlake 1977; Andersen 2001). Also discussed here only in passing is change in morphology and its typological correlates

diachrony

235

(cf. Kiparsky 2008b; McWhorter 2008). We seldom have enough data to be able to see how morphology and syntax interact at the micro-level in language change; whether, for example, word order flexibility precedes development of robust agreement inflection or follows it, or is developed contemporaneously (that is, the same processes that lead to one also facilitate the other). This correlation between morphological form and syntactic parameters is seen most clearly in polysynthetic languages, where there are robust syntactic correlates of extensive morphology. While solving problems like this is beyond the scope of this chapter, it is clearly an important area of morphological change. Finally, I will be paying little attention to morphological change and social constraints. While there has recently been prominent work in the role of cultural constraints on grammaticalization (Evans, N. 2003; Evans and Levinson 2009), and while there is clearly a cultural element to types of grammaticalization (particularly in derivational morphology and the grammaticalization of social categories in language, such as honorifics and ‘kintax’ (Evans, N. 2003)), detailed discussion here is beyond the scope of this chapter. My goal has been to synthesise work in diverse areas of historical linguistics which is of particular relevance for synchronic morphologists.

10.2 Life-cycles of Inf lection: Types of Change in Inf lection ............................................................................................................................................................................. Several previous authors have identified a typology of the changes which take place in morphology. The most comprehensive of these are Koch (1996) and Anderson (1988); their changes include, in addition to regular sound change which operates on fully inflected forms, various types of boundary shift (such as the absorption of material into stems or the reanalysis of one morpheme as two1 ), and analogical changes such as paradigm regularization. Other views of morphological change have focused either on the creation of inflectional material through grammaticalization,2 or on analogical changes (Lahiri 2003). Of the four chapters on morphological change in Joseph and Janda (2003), for example, two deal with analogy, one (Anttila 2003) discusses naturalness, and one focuses on the creation of morphological material from syntax.

10.2.1 Typologies of Morphological Change In this section I present an overview of types of morphological change, drawing heavily from Anderson (1988), Koch (1996), and Chafe (1998, 1999) (see also Haspelmath 1

Changes of this type are described by Haspelmath (1994: 4) under the term ‘telescoping’. See Hopper and Traugott (1993) along with much subsequent work, including Traugott and Heine (1991) and Fischer (2004) 2

236

claire bowern

1994). The typology of boundary changes is due to Koch (1996), while those of morphological interaction are largely accounted for in Chafe (1999). The overall categorization of changes, however, is my own. Morphological change can be divided into four different areas, which are summarized in (1) and discussed in more detail in the following sections. (1)

a. Change in the formal realization of morphemes (allomorphy) (Section 10.2.1.1) b. Change in the placement of boundaries between morphemes (Section 10.2.1.2) c. Creation, loss, and change of morphological categories (Section 10.2.1.3) d. Change in morpheme ordering (Section 10.2.1.4) e. Change in the content, meaning, or function (Section 10.2.1.5)

Note that topics such as ‘analogy’, ‘refunctionalization’, and morphophonology are here not treated as unitary, but are instead spread over several sub-areas of change. This treatment is deliberate. After all, analogical change is not a unitary phenomenon; rather, analogy is a description of a mechanism by which forms which are related in some way interact with each other.3 Furthermore, in providing subtypes of morphological change, I do not mean to imply that more than one type of change cannot happen simultaneously.

10.2.1.1 Changes in Formal Realization Morphology can change through sound change. That is, sound changes apply to fully inflected words, and not to word pieces. Such a statement is unsurprising, of course, since sound change is a consequence of factors involving speech perception and speech production on sound strings. It does not apply directly to roots or affixes within the mental lexicon. Examples are not difficult to find; the example given here in Table 10.1 is from Bethwyn Evans (pers. comm.) and illustrates cognate words in five Oceanic languages (see also Evans, B. 2003: 240–66). The multiplicative prefix can be reconstructed to Proto-Oceanic as *paka-; the changes illustrated here are morphological in that they apply to an affix, but are no different from regular sound changes elsewhere in the lexicon. Sound change can create allomorphy. This occurs whenever an environmentally conditioned sound change occurs on an affix which appears in multiple environments, not all of which are subject to the change. The sound change of the loss of intervocalic *s in Greek, for example, affected stems which ended in a vowel, but not those that ended in a consonant. This created two realizations of the future in Attic Greek— one with s, the other without. Compare the future of a consonant-final root such as deik-s-¯o (δε´ιξω) ‘point out’, which retains the s, with a vowel-final stem such as 3 See also Hock (2003) for discussion of the term, and Anttila (2003) for different ways in which the term can be interpreted.

diachrony

237

Table 10.1 Multiplicative prefix in selected Oceanic languages Language

(be) two

do/happen twice

sound changes the language

Sinaugoro Nakanai Arosi Bauan Fijian Samoan

ruarua i-lua rua rua lua

vaa-ruarua vaka-lua haʔa-rua vaka-rua faʔa-lua

*p > v, *k > ∅ *p > v *p > h, *k > ʔ *p > v *p > f, *k > ʔ

*paka-

phile- ‘love’, which has future philõ (ϕɩλω), ˜ contracted from phile-¯o (ϕɩλ´εω) and earlier from *phile-s-¯o. A similar stem-conditioned creation of allomorphy through sound change is found in the Arandic (Pama-Nyungan) language Kaytetye, where the imperfective suffix has two allormorphs, -rranytye (IPA /-raɲɟə/) and -ranytye (IPA /-ɹaɲɟə/). The form with the trill is the general (elsewhere) allomorph, while the allomorph with the glide is found after stems containing coronals (Koch 1996: 225). Anderson (1988) provides extensive discussion of this type of change and its implications for phonological theory. The previous examples all involved the creation of morphological alternations through sound change, and that is the primary way in which such alternations arise. However, alternations are occasionally created through semantic splits, rather than through phonology. In such cases, the allomorphs are associated with particular meanings, rather than phonological environments. In Proto-Karnic, for example, the ergative case had two allomorphs, *-lu and *-ŋku, which were most probably conditioned by the number of syllables in the stem. In the daughter language PittaPitta (Blake 1979), however, the former allomorph has been reanalysed as the past tense ergative marker, while -ŋu, the descendent of *-ŋku, marks future ergative and nominative subjects (Bowern 2004b). It has occasionally been claimed that morphology can condition sound change; that is, that some morphemes are immune from sound change, or that changes can be confined to certain word classes.4 This is, however, controversial, and many apparent morphologically conditioned sound changes appear to have other plausible explanations once they are further investigated. For example, several subgroups of Pama-Nyungan languages show a rule of homorganic cluster reduction which appears confined to affixes. Wider consideration, however, shows that the rule most likely applies to clusters in the third syllable (or later) in the word. The most common reconstructible clusters in this position come from case markers, however, since longer 4 This view is implicitly entailed by long-distance reconstruction work that compares morphemes of similar shape in the absence of regular sound correspondences in lexical roots.

238

claire bowern

reconstructible roots are quite rare. See Alpher (2004) for further information. Further examples of such changes can be found in Garrett and Blevins (2009), who present an analysis of three cases where morphophonological rules arise from analogical extension rather than sound change. Grammatical ‘trapping’ can also lead to the appearance of morphology conditioning sound change. In such cases, a morpheme or clitic is preserved in some phonological environments but not in others. This leads to the appearance of a grammatically conditioned sound change. Consider the case of l-loss in Northern Paamese (Crowley and Bowern 2010: 174). Crowley presents what is at first sight a fairly clear case of grammatical conditioning of a sound change. In this language, there is a sound change that deletes initial *l from words before non-high vowels, and medial *l between *e and a non-high vowel. The change word-initially does not, however, apply to verbs. Illustration is given in (2): (2)

a. Nouns: i. *leiai > eiai ‘bush’ ii. *alete > aet ‘flat area’ iii. *gela > kea ‘he/she crawled’ b. Verbs: i. *leheie > lehei ‘he/she pulled it’ ii. *loho > loh ‘he/she ran’

Blevins and Lynch (2009) show that the conditioning environment is explicable without an appeal to grammatically conditioned sound change. They argue that the third person singular irrealis prefix *mi- (and other high-vowel prefixes) inhibited the application of the sound change. However, those prefixes were subsequently lost. Analogical changes provide another type of change in formal morpheme realization; some types of analogies result in the replacement of one morphological form with another. Such changes are well-known from the handbooks and need not be rehearsed here. Sound changes can be subject to secondary analogical changes which obscure their origin in regular sound change. For example, in the Greek example described above, several classes of vowel-final roots have had the s-allomorph of the future reintroduced by analogy to forms that would have regularly lost it.

10.2.1.2 Boundary placement changes The second type of morphological change involves the placement of morpheme boundaries (or, in non-morphemic theories of word formation, on changes in word formation rules). Boundaries may be lost, created, or moved within a word. This may occur through several different processes, including analogical changes which result in the reanalysis of the internal structure of a word, thus changing the placement of the morpheme boundary. This is found, for example, in a number of words borrowed into English which end in s, including cerise ‘cherries’ (most probably from Old Northern

diachrony

239

French cherise). In this case, the word cheris(e) was reanalysed as plural and the final consonant (which was part of the root) was taken as the plural marker. Such cases of boundary reanalysis are no doubt aided by the fact that cherries tend to occur in clusters, thus making plural forms for this word rather more frequent than for items which tend to occur singly. It remains to be seen whether there is similar frequency skewing which biases boundary shifts in other cases. Boundaries may be lost when productive morphology becomes unproductive. This is more frequently associated with derivational morphology than inflectional, but examples of inflectional morphology being absorbed into stems do exist. One common case involves the fossilization of case marking in adverbial words; several examples are found in Bardi (Nyulnyulan), where the ‘source’ case -joon appears fossilized in adverbs, and in the occasional noun such as loomijoon ‘orphan’. The word loomijoon is no longer treated as derived from loomi and can receive additional case marking. Latin also shows boundary deletion in adverbs such as qu¯a ‘where, in which place’, where the form, although historically a feminine ablative singular, no longer declines and no longer has the properties of an ablative-marked nominal. A common cause of morpheme boundary movement involves a change where a sound change triggers a reanalysis. The following example is from Longgu, but equivalent cases are found in many other parts of the world. In some cases, stems ending in a consonant lose that consonant at a word boundary; when the stem is followed by a suffix, however, the consonant resurfaces; a sample paradigm is given in Table 10.2 (forms from Hill 1992).5 In the case of Longgu, transitive verb forms are built from intransitive ones by the addition of a suffix; this suffix was historically *-i (see Evans, B. 2003) and the first consonant of the synchronic transitive marker was part of the root. A sound change which led to the loss of final consonants in the language also led to the reanalysis of the form of the transitive morpheme as -Ci. Since the final root consonants were unpredictable in pre-Longgu, this change has led to arbitrary conjugation classes in the language.

Table 10.2 Longgu stem alternations Intrans.

Gloss

Transitive

Gloss

Suffix allomorph

bere aŋi eno maʔu moa lilimo

look cry lie down be frightened vomit drown

bere-ŋiaŋi-sieno-vimaʔu-nimoa-li limo-zi-

see sth cry for sth lie on sth be frightened of sth vomit on sth drown s.o.

-ŋi -si -vi -ni -li -zi

5 Another common pattern of this type is when vowels are lost in word-final position, but resurface when further material follows.

240

claire bowern

A similar change has led to unpredictable forms in M¯aori passive formation (Biggs 1961; Hale 1974; Kenstowicz and Kisseberth 1979).6 In other cases, there is no such sound change to trigger the reananalysis. Haspelmath (1994: 9) describes a case of boundary shift which he calls ‘root secretion’. In this change, a non-affixal part of a root is reanalysed as part of a suffix, which is then generalized to other words. Haspelmath gives the example of Greek mélan- ‘black’ and its causative melaín-(¯o) ‘make black’. The causative is reanalysed as -aín- and extended to other words, such as leuk-aín-(¯o) ‘make white’ (root leuk-). Most of Haspelmath’s examples come from derivational morphology, but such changes are also sporadically found in inflectional material. Vondrák (1906: 11) gives the example of Slavonic, where the final vowel of u- and i- stems was reanalysed as a case ending. The ending subsequently spread to other declension classes. As Haspelmath points out (and before him, Plank 1981: 75), changes of this type provide evidence that the model of word formation which most parsimoniously accounts for diachrony in morphology is one which does not appeal to strict segmentation. A further example of boundary shifts is known as Watkins’ Law. This name was given by Koch (1994) to the process by which third person singular marked forms are reanalysed as containing no affixal material; they then serve as the base for other inflections. This is illustrated by the changes in the Old French paradigm given in Table 10.3. While such changes are well described in the literature, the triggers for such boundary reanalyses are not, as yet, well understood. Some general principles can be noted, and some specific rationales for why some changes occur more frequently than others can be proposed. Some such changes are presumably driven primarily by acquisition. Others, such as Watkins’ Law, are perhaps more puzzling, or are shaped by more subtle

Table 10.3 Old French reanalysis of the third person singular as unmarked

1sg 2sg 3sg 1pl 2pl 3pl

Earlier

Later

canté-i canté-st canté-t canté-m canté-tz canté-ren

cantét-e cantét-es cantét-∅ cantét-em cantét-etz cantét-on

6 In cases like these, it can be difficult to tell whether the rule is best treated as a historical sound change or as a synchronic phonological rule which had its origins in sound change. Evidence that the historical final consonant has been reanalysed as a conjugation marker comes from sources such as unetymological final consonants (that is, from roots which show alternation but where the final consonant is not predicted from sound change) and from the treatment of roots newly introduced to the language.

diachrony

241

processes. For example, while linguists note a cross-linguistic tendency for third person forms to be bare roots or stems, speakers do not, of course, know this, and in the Old French case, there are no morphological cues within the language that would lead speakers to an analysis of third person subject verbs as null marked. Some such changes are discussed under the rubric of ‘natural morphology’ (Dressler 2003).

10.2.1.3 Creation, loss, and change of morphological categories The changes discussed thus far have been system-preserving in that no new morphological categories have been created or lost. Many morphological changes are indeed of this type, for categories may be reinforced with new morphological material (see, for example Heath 1997, 1998) if eroded. This apparent stability has been the rationale for the use of morphological categories—without necessary correspondences in form—in language classification (see further Section 10.4). Hymes (1956), for example, used morphological categories and their relative orders to argue for a relationship between Athabascan languages, Haida and Tlingit, even in the absence of cognate morphemes. Others invoke a concept of morphological stability in more general terms. Haspelmath (1993: 306), for example, argues that ‘certain types of change are motivated by speakers’ attempts not to deviate too much from older patterns.’ Not all morphological features are equally stable, however. Moreover, there is substantial disagreement regarding which features are likely to be typologically persistent, and therefore diachronically stable. Proposals are given in Nichols (1992) and Wichmann and Holman (2011), amongst others. On current data, however, it is very difficult to systematically distinguish temporally stable features from those which are easily diffusable. In the former case, the feature is old and persistent; in the latter, it is widespread, but may or may not be stable once acquired.7 Heine and Kuteva (2005) and Heine (2011) have argued for the recognition of a concept of ‘grammaticalization area’; that is, a set of contact-induced changes amongst languages in contact in a particular geographical area where languages grammaticalize the same features. Heine (2011) provides examples of parallel grammaticalization under language contact. The presence of such areas is not unexpected if there is a heavy social component to language change; however, more explicit discussion is needed about the assumptions that identification of ‘grammaticalization areas’ entail. New categories in a language can be the result of extensive language contact. Ross (2001) gives numerous examples of syntactic and morphological features which have diffused between Takia and Waskia. Adelaar (2006: 300–1) gives the examples of the distributive category marked by a suffix -kama, which has been borrowed from 7 As an example of the problem of distinguishing the two, consider the presence of contrastive tone in a language. Authors such as Dediu (2011) have considered tone to be a very stable feature. However, we know from studies of contact phenomena in South East Asia (e.g. Enfield 2006, 2005) that tone systems are highly diffusable, and so their presence in a large number of languages from different families in a region does not make the feature reconstructible.

242

claire bowern

Quechua into Amuesha. Bardi (Nyulnyulan, Australian) has an adjectival suffix -wan, which comes ultimately from English ‘one’, via North Australian Kriol. Although much of this section has focused on the role of language contact in the creation of new morphological categories, this is not the only way in which we find new categories arising. One way is through the routinization and grammaticalization of material that exist in collocations in earlier stages of the language. Evans, N. (2003: 23–9) provides discussion of the grammaticalization of kinship marking in pronouns or verbal marking in several unrelated languages in Australia, and Chafe (1999) invokes a similar principle to explain the creation of new pronoun forms in Iroquian languages. Grammaticalization processes are very well studied, of course; see Heine and Kuteva (2005) and Hopper and Traugott (1993) for overviews, amongst many other references. However, grammaticalization does not only create new morphological categories. Such processes may also add members to existing paradigms. In the history of Latin, for example, the imperfect endings -bam, -bas, -bat, etc, were grammaticalized from periphrastic auxiliary constructions comprising the main verb and a conjugated form of the verb ‘to be’ (see further Sihler 2008: 554–5). Finally, morphological categories may change through the amalgamation or split of paradigms. Newman (1980), for example, describes a number of such changes in Salish languages. In this type of change, paradigms are amalgamated, syncretised, or otherwise blended to form new paradigms.

10.2.1.4 Changes in the order of morphemes Morphological reordering as a raw change is not common, although changes which result in permutations of the underlying (linear) order of affixes are found in language. This often, though not always, proceeds from change from less iconic to more iconic ordering; for example, derivational inflection closer to the root, or case affixes outside of number affixes. In other cases, an apparent shift in morpheme order is the result of a more complex set of changes. Gur languages, for example, have noun class affixes, just as most Bantu languages do. However, the Gur affixes are suffixes, while languages elsewhere in Niger-Congo (including Bantu) have prefixes. The material marking the noun classes is cognate, it simply occurs in a different position in the word (Dimmendaal 2001a: 377–9). Prefixation can be reconstructed to Proto-Niger-Congo (Williamson 1989: 31–40), and the sequences of change which led to suffixation in Gur were as follows. Bantu languages have widespread noun class concord within the Noun Phrase; this concord is marked on demonstratives, for example. Williamson argues that demonstratives were post-nominal (as they are in contemporary Bantu languages). The demonstratives were semantically bleached, phonologically reduced, and subsequently grammaticalized as suffixes to the noun. The earlier noun prefix markers were then lost through sound change, leaving formally similar (and cognate) material as suffixes. In this case, the morpheme reordering is an incidental outcome of a chain of changes affecting word structure, phonology, and morphology.

diachrony

243

Morphological categories may be reinforced with additional marking, leading to change. This was seen in the Gur example in the previous paragraph, although the reinforcement was subsequently lost. A clear example of reinforcement in morphology comes from Donohue’s (2003) analysis of the origin of complex subject agreement in the Skou language of northern New Guinea. In this language, agreement for subject is marked in several different ways—by a proclitic, which can appear on both the subject NP and the verb; the verb itself may take both proclitic subject markers and prefixes which indicate person and number. Finally, the verb itself shows alternations in vowels for features of the subject, such as whether it is plural, or feminine (see Donohue 2003: 481–7). An example is given in (3): (3)

a. Nì nì=lúe 1sg 1sg=1sg-know ‘I know.’ b. Mè mè=p-úe 2sg 2sg-=2sg-know ‘You (sg) know.’ c. Te te=r-í 3pl 3pl=3pl-know.pl ‘They know.’

(All examples from Donohue 2003: 485, ex. (22))

The word lúe hear, know is part of a paradigm which shows both initial consonant variation for subject person and number, while the vowel of the verb also co-varies with features of the subject (rí for third person plural, for example). Donohue shows from comparative evidence that a probable trigger for this marking reinforcement was a sound change that led to the reduction of consonant clusters. He argues that as sound change has eroded the morphological marking of the agreement categories, new markers have been added to reinforce the distinctions. However, we have no way of telling when such categories will be reinforced (as in Skou), or when they simply disappear from the language.

10.2.1.5 Changes in meaning The final area of change here is semantic change. While we saw in Section 10.2.1.1 that morphology does not interact with sound change independently from the stem to which it attaches, in the domain of semantic change, we do have some evidence for the autonomy of morphology. That is, morphological categories can change meaning independently of the stems to which those morphemes attach. In other cases, fully inflected words can change their meaning, leading to morpheme boundary shifts and creation of sub-paradigms. Chafe (1999, 2008) illustrates this process using examples from Seneca. Finally, some changes which appear to be confined to morphology alone are the reflexes of changes within larger structures. Many instances of semantic change in morphology involve the same processes of change that we see in other domains of semantics. For example, meaning can be

244

claire bowern

broadened or narrowed, and there can be changes in meaning driven by metaphor or metonymy. Some of these changes appear to be unidirectional. Semantic change in progressive verbal tense markers, for example, shows an overwhelming tendency towards bleaching rather than narrowing (Deo 2006). Some semantic change in morphology can lead to very large changes in meaning. As mentioned above, in the Australian language Pitta-Pitta (Karnic, Pama-Nyungan), for example, certain case markers have been reanalysed as tense markers (see Bowern 2004a for details). The ergative allomorphs -ŋu and -lu have split, with the former marking the subjects (both transitive and intransitive) of future verbs; the latter is the ergative marker in non-future clauses. The trigger for the change in the morphology was most likely not, however, within the case marking system directly, but a more general reanalysis of a construction marking unrealized actions, involving both case marking and verbal morphology. There are other instances of change, however, where change in morphological function does not appear to be tied to a larger construction, but is confined to the morpheme directly. We find, for example, shifts in the marking of spatial cases. In the history of Latin, for example, the ablative case absorbed the functions previously associated with the locative, which fell out of use except for a few fossilized forms. How such processes occur in grammar remains controversial. For those who model such changes at the level of individual grammars, the problem can be avoided somewhat by treating the object of analysis as two synchronic grammars, each inferred by each speaker from the surrounding data and therefore only indirectly comparable (Hale 2007). For those who take a population view of language change, however, the preconditions of change and their mechanics require more detailed investigation (see further Urban 2011 for discussion of these points). For example, it is unclear whether such changes occur to ‘constructions’ or to specific lexemes (or lexical collocations). Analogies may be drawn with sound change here, where it is debated whether changes operate on individual lexical items and subsequently spread through analogy, or whether they operate on sounds and sound classes. Labov (2007) and other work provides evidence that both types of changes may occur under different conditions. The same may well be true of morphological change. Chafe’s work on change and the rise and reinforcement of morphological complexity (e.g. Chafe 1998, 1999) emphasizes processes which are both analogical and which originate in particular lexical items rather than particular lexical categories. For example, he includes examples of sporadic paradigmatic gaps, ad hoc grammaticalization and paradigm split based on single lexical items. However, he also has examples of changes based on stem shape (Chafe 1998: 109), which would be an example of categorial change. A final type of semantic change within morphology is known as exaptation (Lass 1990) or refunctionalization (Smith 2008). This is the process whereby one morphologically marked opposition is recruited to mark a different type of opposition. That is, it is not a change in formal marking per se, since the marking remains; rather, the same morphological markers come to mark a different pair of oppositions. Smith (2008) describes the case of Young People’s Tiwi (Lee 1987), where a distinction between first person inclusive and exclusive has been remapped as a past/non-past tense distinction.

diachrony

245

Lass (1990) provides examples from Germanic and Smith (2006) from Romance. Smith (2008) has argued that in such cases, the remapping of oppositions is not arbitrary, but rather follows a more general principle of ‘core-to-core’ mapping, whereby oppositions retain their markedness relations to one another, even if the domain of marking changes (e.g. from person to tense, as in the Tiwi example).

10.2.2 Morphology as a Shaper of Change The idea that there is autonomy in morphological structure has taken several forms. Synchronically, there is Aronoff ’s work (e.g. Aronoff 1994), based around the concept of a ‘morphome’ or abstract layer which mediates form and function (analogous to the phoneme in phonology).8 Maiden (1992, 2005, 2009) provides a diachronic view of morphological autonomy in which certain structural features in language are selfperpetuating; that is, that there are certain structures which lead to either the repetition of a particular change over many generations (Kastovsky 2006), or (conversely) that certain structures (paradigmatic shapes, for example) both constrain and drive change (Maiden 2009). In the former case, we see a predispositon in the structures of a language for speakers to (re)analyse parts of the morphology in certain ways. In the latter case, there is the idea that paradigmatic shapes may be maintained, independent of the particular forms within the paradigm and independent of other changes; this is a diachronic interpretation of Aronoff ’s synchronic view. An alternative (and broader) statement of this view is Haspelmath’s (1993); Haspelmath comments that some oppositions in morphology may be stable irrespective of the forms that mark them. Such ideas are not widely accepted in historical morphology. Anderson (2011), for example, provides a detailed critique of Maiden’s analysis of Rumantsch data and argues that the patterns which argue for autonomous morphology can also be described by straightforward phonological conditioning. A further curious paradox is that discussions of morphological autonomy have tended to stress its stability; even while much work on morphological instability has appeared (particularly under the study of analogical changes).

10.3 Sources of Morphological Inf lection and Morphological Clines ............................................................................................................................................................................. New morphology arises within a language from either the grammaticalization of independent lexical items, or changes within the morphological system (as described in Section 10.2). That is, new morphological forms can be created either from within the existing system, or by the introduction of new material from elsewhere in the lexicon. 8 See Round (2009) for a development of this idea and its consequences for Kayardild (Tangkic, Australia).

246

claire bowern

10.3.1 Grammaticalization and the Pathways of Change For many types of inflection, pathways of morphological sources are clearly identifiable. Tense markers, for example, often have their sources in independent verbs, serial verb constructions, or auxiliary verbs. The Latin imperfect, for example, is a compound tense built around the univerbation of the verb stem and the verb ‘be’. An alternative path for tense marking comes from the interaction between tense marking and nominalization. Epps (2008) provides an example from Hup (Nada-Hup, Amazonia) where the word tég ‘wood’ has been grammaticalized as a generic marker, which in turn can be applied to verbs as a purposive and in future constructions.9 The Hup example illustrates a grammaticalization cline—a chain of changes operating on forms which may carry them grammatically and semantically far from their origins. Parts of a cline have very circumscribed origins, while others may be more diffuse. For example, person agreement marking appears to have only two origins. By far the most common is pronominal affixation. That is, in the vast majority of cases where the etymology of the agreement markers is known, argument agreement arises through the univerbation of personal pronouns with the verb. In rare cases, individual agreement forms also arise from shifts in morphological boundaries or templates. For example, the Warrwa third person plural subject marker ŋ- originates in Nyulnyulan as a past tense marker, but the loss of the third person prefix * i- in this language leads to a reanalysis of the past marker as the desinence of third person in plural forms. Thus sources of agreement are highly constrained. The etymological sources of free pronouns, however, are rather broader. They include words for ‘body’, demonstratives, and both honorific and pejorative phrases.

10.3.2 Morphological Cycles Morphological cycles were an important part of early work on morphology. The idea that languages may move from isolating type to agglutinating through univerbation, and to inflectional through sound change, originates with Humboldt (1836), though both the typology of morphological type and the relations between types has been the subject of extensive debate (from Sapir 1921 to Haspelmath 2009. Much recent work (particularly in grammaticalization theory) has de-emphasized the role of such cycles in favour of work which derives the cycles from other processes.10 Grammaticalization theory predicts the accretion of grammatical material; sound change leads to obscurity in morphological boundaries; other processes lead to the creation of irregularity in 9

Heine and Kuteva (2005) is a compendium of pathways. For an exception to this view, see Dixon’s (2002) description of morphological change in Australian languages, which treats such cycles as a property of these languages, in the manner of an ‘invisible hand’ process as described in Keller (1989). Such a view is not widely accepted amongst Australianists, however. 10

diachrony

247

paradigms (and thus to inflectional and fusional languages). Conversely, sound change may also lead to extensive erosion in morphology. However, there are two areas of morphological cycles where recent work has challenged a purely cyclic view. The first is the way in which extensive morphology is lost; the second is in the creation of polysynthesis. Earlier views of lack of morphology either treated the cycle as a single process (that is, all languages started without morphology and came to acquire it), or that sound change was the only mechanism which led to the loss of morphological categories. However, as we have seen, and as work by Keller (1989), Donohue (2003) and others have shown, there are processes which frequently lead to a continual reinforcement of morphology when categories are eroded through sound change. A major other cause of morphological loss is language contact, particularly the situation where large numbers of second language speakers adopt the new language, but without the morphology. Another case, also well documented in the literature, is during the last few generations of language death. In such cases, speakers may acquire the lexical items of the language, but do not acquire the morphological complexity.11 Thus morphology can be lost, both as languages lose speakers and as they acquire them. A second area of problems in a cyclic view of morphology is in the treatment of polysynthetic morphology. Since the work of Hale (1983), Jelinek (1984), Baker (1988), Austin and Bresnan (1996), and others, it is increasingly the case that polysynthetic morphology is treated not just as a property of the morphology, but as a correlate of a set of syntactic features which interact with morphology. (Of course, we have not discussed morphology–syntax interactions in other parts of morphology, and to leave them out in the other cases but to discuss them here is to paint a skewed picture of language, but nonetheless, the syntactic variation amongst the languages described as polysynthetic is much less than amongst those of other morphological types. See Julien 2002 for an overview.) Thus there does not seem to be much evidence for a direct change from inflectional to polysynthetic morphology, within morphology, but from syntactic changes and syntactic properties which shape the morphology. van Gelderen and Willie’s (2011) work on Athabascan polysynthesis in historical perspective illustrates this view.

10.4 Inf lectional Morphology in Language Classif ication ............................................................................................................................................................................. Inflectional morphology has played an important role in language classification, both because of its relative stability (compared to sound changes) and relative lack of 11 For examples in Australia, see Lee (1987) for Tiwi and Scheffler (1978) for Dyirbal, amongst others. It should be noted that extreme morphological loss is not the only outcome of this type of language contact, as O’Shannessy (2005) and McConvell and Meakins (2005) in particular, have shown.

248

claire bowern

borrowability (compared to lexical items). Such evidence is weighted differently by specialists in different families, however. In some families, linguists working on the internal subgrouping of languages have tended to privilege the evidence of morphological correspondences over that from shared lexicon or sound change. Such has been the case in Australia, for example, where perceived high rates of borrowing, combined with relatively small amounts of sound change (and few highly distinctive changes which would provide valuable evidence in subgrouping), have led linguists to base much of their classificatory work on correspondences in morphological form (for further comments see Bowern and Koch 2004; Koch 2004). Another argument for the use of inflectional morphology is based on morphological stability. Because morphology tends to be less subject to replacement than lexical items, it allows access to potential genetic relationships between languages where the ancestor is more temporally remote than that which can be reconstructed using the comparative method. This has been the case with several non-Pama-Nyungan language families in Northern Australia; see, for example, Green (2003) for the demonstration of an ‘Arnhem’ group in Central Northern Australia. Such families raise questions, however. It is not at all clear what processes give rise to languages which may have very similar phoneme inventories, similar morphological inventories, but very low rates of shared lexicon. One possibility is contact-induced transfer (see Ross 2007); another is very high rates of lexical replacement caused through any number of factors, such as strong necronym taboos. The most famous use of this argument comes from Hymes (1956) and his work on the relationships between Athabaskan, Tlingit, and Haida. Hymes (1956: 631) compares the order and semantics of morphology in these languages to both argue for a relationship between the three and to reconstruct the positional order for ProtoNa-Déné. Hymes’s work has remained controversial, however, because he compared only category positions, and not forms, and because there are universal tendencies in morphological ordering (see Julien 2002 and Comrie 1989 for a summary). Furthermore, it is difficult to reconcile the claims of Hymes with those from Donohue (2003) and other work on morphological accretion, which shows that morphological categories may be renewed, but in different places in the word. Finally, we have examples where the morphological template can change, even when the forms themselves are stable (see Bowern 2012 for discussion and examples). Unfortunately, at this time we lack the detailed quantitative work that would allow us to assess morphological stability across a range of families. We are unable to assess whether such similarities are most likely due to chance, or to universal tendencies. The comparison of position, forms, and lexical items along with sound change reconstruction obviates this problem. The insistence on formal correspondences along with functional and positional ones when establishing language relationships is not controversial (Campbell and Poser 2008), although as mentioned in Section 10.2.1.3 above, there is still plenty of recent work in this area which relaxes this constraint.

diachrony

249

10.5 Conclusions

.............................................................................................................................................................................

In summary, I have provided an overview of changes in morphology, life-cycles of change, and identified areas in which morphological change is triggered from elsewhere in the language. On close examination, many of the case studies which explain change through morphological autonomy are also explicable through other processes; there appears to be very little evidence for change which operates on morphology alone. Morphological forms interact with semantics, syntax, and phonology, often leading to changes whose origins lie outside of purely morphological processes such as analogy. Even changes which operate purely on morphological class, such as paradigm reorganization, may have a semantic or syntactic origin. Basic facts of the diachrony of inflectional morphology have not altered very much in the last twenty years. For example, though we now have a much greater knowledge of the variety of morphological changes which occur in natural language, the types of change that have been observed have not changed a great deal.12 Where debate has focused, however, is on the nature of explanations of change in morphology, the relationship between language contact and morphological change, and causes of change which have their origins outside the linguistic system. Debate continues on stability in morphology; while on the one hand we can see a large body of work which catalogues the changes which take place in both morphemes and morphological systems, on the other hand, there is still a widespread view that stability of inflectional systems make them a good data source for identifying remote genetic relationships.

12

One exception is the description of endoclisis; see Harris (2000).

chapter 11 ........................................................................................................

CONTACT-INDUCED CHANGE ........................................................................................................

maarten kossmann

11.1 Basic Introduction

.............................................................................................................................................................................

Inflectional systems change over time. This change is due both to internal factors and to the influence of other languages. This chapter will draw the main lines along which inflectional forms and patterns are transferred from one language to another. There are many ways to tackle the question why and how this change takes place. Here I shall focus on two points: First, are the elements taken over as isolated forms (e.g. transfer of a morpheme) or are they taken over as part of something else (e.g. transfer of morphology together with lexemes)? Second, to what extent do the transferred elements concern new categories in the language and to what extent do they just constitute formal additions or substitutions in a system that otherwise remains the same? Of course, other ways of tackling the subject are possible, and have been chosen in the past. Thus basic texts such as Thomason and Kaufman (1988) and Van Coetsem (1988, 2000) take the sociolinguistic situation of the transfer as a point of departure, while other researchers focus on cognitive and other underlying pressures that precondition contact-induced change (e.g. Matras 2009). I shall not go into these approaches in this article, although the reader will recognize my indebtedness to them at certain points. As stated above, the focus in this article is on the transfer of forms and patterns. It therefore excludes another type of linguistic change in which the presence of nonnative varieties plays a role (Breu 1996): large-scale simplification of morphology (Kusters 2003), culminating in the morphological stripping typical of pidgin and creole languages (Siegel 2007; Crowley 2008; and the principled discussion in Owens 2001). Contact-induced change falls into two basic types. In the first type, the introduction of foreign morphology is concomitant to the introduction of other materials, almost always lexical in nature. In the second type, morphological forms and patterns are

252

maarten kossmann

taken over as isolated elements, that is, they are not concomitant with the take-over of other elements.

11.2 Concomitant Borrowing

.............................................................................................................................................................................

Inflectional elements can be borrowed in connection with other elements, mainly lexical items. When only one form of an inflectional paradigm is taken over, this does not lead to any changes in the morphology of the recipient language: the former inflectional element functions as part of the lexical item. Thus there is no need to consider the final sequence -um in French album as anything but part of the French stem, even though in the donor language, Latin, it is a marker of number, case, and gender. However, in many languages several inflectional forms are taken over with one single lexical item. In such cases, the oppositional value is retained, and the inflections still have morphological value. This is found, for example, in English plurals on a Latin or Greek base, such as sg phenomen-on pl phenomen-a. As these forms function within a larger system, which also has native inflectional forms (e.g. in English sg simpleton pl simpleton-s), the borrowed forms constitute a morphological subsystem on their own, parallel to the native paradigms. In Kossmann (2010, 2013), this phenomenon has been christened Parallel System Borrowing. Parallel System Borrowing is well-known from learned contexts, especially with nouns, in European languages (mainly Latin and Greek forms) and in languages of the Muslim world (mainly Arabic forms). However, it is not restricted to learned contexts, nor is it restricted to nominal morphology. One language which has elaborate inflectional parallel systems is Ghomara Berber, a variety of Berber spoken in northwestern Morocco (El Hannouche 2008, Mourigh 2015). Virtually all inflectional paradigms show parallel systems with dialectal Arabic forms. In addition to plurality marking on nouns, one finds parallel systems in adjectival agreement, in bound pronominal forms, and in verbal inflection. Table 11.1 illustrates the parallel paradigms of adjectival agreement in a system that expresses gender (masculine vs. feminine) and number (singular vs. plural). The number of Berber-based adjectives is very low, and the

Table 11.1 Adjectives in Ghomaran Berber

sg.m sg.f pl

Berber adjectives

Arabic-based adjectives

-∅ -əθ -əθ

-∅ -a -in

məllul məllut-əθ məllul-əθ ‘white’

Source: Mourigh (2015).

rqiq rqiq-a rqiq-in ‘slim’

contact-induced change

253

Table 11.2 Verbs in Ghomaran Berber Berber paradigm Ghomara

Perfective

Imperfective

Future

1sg 2sg 3sg.m 3sg.f

z.z.a-x t-əz.z.a-t i-z.z.a t-əz.z.a

təz.z.a-x əh-təz.z.a-t i-təz.z.a h-təz.z.a

(š) a=z.z.u-x (š) a=z.z.u-t (š) a=z.z.u (š) a=z.z.u

1pl 2pl 3pl

n-əz.z.a t-əz.z.a-m z.z.a-n ‘to plant’

n-təz.z.a h-təz.z.a-m təz.z.a-n

(š) a=n-əz.z.u (š) a=z.z.u-m (š) a=z.z.u-n

Ghomara

Perfective

Imperfective

Future

1sg 2sg 3sg.m 3sg.f

.s.saði-θ .s.saði-θi .s.sað .s.sað-əθ

ka=n-əs..sað ka=d-əs..sað ka=y-əs..sað ka=d-əs..sað

(š) n-əs..sað (ž) d-əs..sað (š) i-s..sað (ž) d-əs..sað

1pl 2pl 3pl

.s.saði-na .s.saði-θum .s.sað-u ‘to fish’

ka=n-əs..sað-u ka=d-əs..sað-u ka=y-əs..sað-u

(š) n-əs..sað-u (ž) d-əs..sað-u (š) i-s..sað-u

Arabic paradigm

Source: Mourigh (2015).

Arabic-based morphology can be considered the ‘regular’ way of making an adjective in Ghomaran. With bound pronouns, parallel systems appear in connection with particles that have been borrowed from Arabic (to what extent this should be considered inflection is difficult to say, though). Thus, while Berber-based elements, such as the basic prepositions, are followed by Berber pronominal suffixes, a number of particles borrowed from Arabic are followed by Arabic pronominal suffixes. Parallel systems are also found in verbal inflection. A significant number of verbs borrowed from Arabic retain their Arabic inflection in Ghomaran Berber (Mourigh 2015) Arabic paradigms are found with some high-frequency verbs, which constitute the only option for denoting the activity, such as tlaqa ‘to meet’ or the above example in Table 11.1 .s.sað ‘to fish’ (the data were recorded in a fishing village). Therefore they cannot be considered cases of code switching. The presence of similar forms in several sub-dialects of Ghomaran (cf. El Hannouche 2008 describing another variety of Ghomaran) and their presence in texts taken down in the 1920s (Colin 1929) shows their geographical and historical continuity.

254

maarten kossmann

In fact, with verbs the parallel system reaches outside the inflectional domain, as Arabic clitic object pronouns are used with verbs with Arabic inflection, while Berber clitic object pronouns are used with verbs with Berber inflection. Parallel systems outside the nominal domain are also attested elsewhere; for verbal inflection one may cite a number of Romani cases, the best described of which is Ajia Varvara Romani, which uses Indo-Aryan verb inflections with native verbs, and Turkish verb inflections with verbs of Turkish origin (Igla 1996). One notes that the Romani group speaking this variety was resettled from Anatolia to Greece in the 1920s. The younger generation of this group does not know any Turkish any more, but still uses the Turkish parallel system in their language. Concomitant borrowing often creates a perfect parallel to native morphology, for example in the case of English singular–plural pairs; in such cases there is no impact of the borrowing on the system of categories expressed. It is not unreasonable to assume that the generalized parallelism of inflectional morphology in Ghomara Berber was made easier by the strong parallelism in categories that obtains between Moroccan Arabic and Berber in general, the result of common heritage and over a thousand years of convergent developments.1 In some cases, however, the borrowed inflection functions differently from the native inflection, thus creating morphosyntactic or categorial subsystems. A case of categorical mismatch between the two systems may be found in Cypriot Arabic, which, according to Newton (1964, cf. Kossmann 2008), has case marking in Greek loanwords, but not in native Arabic nouns. More subtly, there can be differences in morphosyntactic behaviour. Thus in Tasawaq, a Northern Songhay language spoken in Niger with elaborate parallel system borrowing from Tuareg (Berber), the system of plural marking is different for native Songhay nouns and nouns of Tuareg origin. With Songhay nouns, an NP-final plural marker is used, which only follows the noun when it is not followed by any other part of the NP. With Tuareg nouns, plural marking takes place on the noun, whatever its position in the NP. The Songhay plural marker cannot follow the Tuareg noun immediately, but may be present when the noun is not the final element of the NP (see Table 11.3; Kossmann 2007: 78). Once a language possesses parallel systems, the borrowed elements may spread outside their etymologically defined domain and appear on words where they do not belong historically. In the first place, this affects foreign vocabulary. In Tasawaq, which has elaborate borrowing of Tuareg parallel systems, among other functions in the formation of noun plurals, a Tuareg ending is used to make plurals of loanwords from Hausa (Kossmann 2007). Similarly, pluralization of Spanish loans in Ghomara Berber is mostly according to Arabic morphology, although a few retain their Spanish plurals (Mourigh 2015).

1 There are a few categories which function differently in (Ghomaran) Berber and in Moroccan Arabic. The way this systemic discrepancy is solved within the parallel systems is described in detail in Mourigh (2015).

contact-induced change

255

Table 11.3 Plural marking in Tasawaq Tasawaq

Songhay noun dábdè dàbdá-yo dàbdè sídày-yo

Tasawaq

‘piece of clothing’ ‘clothes’ (cloth-pl) ‘red clothes’ (cloth red-pl)

Tuareg borrowing tákàrdè sìkárdààwàn sìkárdààwàn sídày-yo

‘sheet (of paper)’ (paper.sg) ‘sheets (of paper)’ (paper.pl) ‘red sheets (of paper)’ (paper.pl red-pl)

Spread may also affect native lexicon. Thus, Tasawaq has a small number of native Songhay words which receive Tuareg plural markers, while in Yiddish the Hebrew m:pl suffix -im is also used in Germanic words such as nar—naronim ‘fool’ and poyər— poyər-im ‘farmer’ (cf. Katz 1987: 57). It should be stressed, however, that this spread is by no means a case of contact-induced change, that is, change related to the presence of a foreign language. Once parallel systems exist in a language, they are part of that language’s system, and can therefore spread just like other morphemes. Their historical background in language contact does not make the process different from other types of analogical spread. A special case of parallel system borrowing is found in Michif, a mixed language with French nominal morphology and Cree verbal morphology (Bakker 1997). Whatever the history behind this compartmentalization, one may assume that it evolved as parallel system borrowing, followed by partial relexification (only involving nouns). The effect of this was a language in which not only all nouns are lexically French, but also their morphologies. The development of parallel systems is almost invariably concomitant with the borrowing of lexicon. There is one example, though, where the parallel system is con˙ comitant to the borrowing of a suffix. In Sebjan-Küöl Even, a Tungusic language of Siberia, two modal suffixes have been taken over from Sakha (aka Yakut), a Turkic language (Pakendorf 2009). In addition to the modal suffixes themselves, the person– number inflections following these suffixes have also been taken over. The first suffix is the so-called presumptive suffix of Sakha, which has presumptive (‘probably’) and ˙ assertive meaning. In Sebjan-Küöl Even, this suffix is taken over as an assertive marker, together with the Sakha person–number affixes. The same has happened with another Sakha modal suffix, the necessitative. An interesting complication is that, in Sakha as ˙ well as in the borrowed Sebjan-Küöl Even forms, the sets of person–number marking are different between the two modal forms, as the presumptive is constructed with Sakha possessive pronouns, and the necessitative with Sakha subject agreement markers. Pakendorf (2009) provides the sample paradigms shown in Table 11.4 (blanks are forms unattested in Pakendorf ’s materials).

256

maarten kossmann ˙ Table 11.4 Sebjan-Küöl Even assertive (presumptive) paradigm of a ːŋŋa ‘spend the night’ and necessitative paradigm of em ‘come’ (Sakha elements are underlined) Assertive 1sg 2sg 3sg

a: ŋŋa-j-dag-i.m

1pl 2pl 3pl

a: ŋŋa-j-dak-pi.t

a: ŋŋa-j-dag-a

a: ŋŋa-j-dak-tara

Necessitative

em-e-jekte:k-kin em-e-jekte:k em-e-jekte:k-pit em-e-jekte:k-kit em-e-jekte:k-e-l

Source: Pakendorf (2009).

Similar constructions, although different in detail, are found in related Uˇcur Evenki (Myreeva 1964: 51, cited in Pakendorf 2009: 88). The two borrowed morphemes which trigger this parallel system represent categories which were (presumably) not expressed in Tungusic morphology before. Thus, this seems to constitute a rare case2 of a parallel system linked to the borrowing of an inflectional morpheme rather than a lexical item.

11.3 Borrowing of Isolated Material

.............................................................................................................................................................................

The borrowing of isolated inflectional material (cf. Gardani 2008) will be presented here in the following way. Borrowing of inflectional forms and patterns presupposes a fairly good knowledge of the contact language by the speakers of the borrowing language. Using two languages implies the problem of categorial dissimilarity: it is often the case that certain categories are absent or different in the two languages involved. When studying the borrowing of patterns and forms, there is a fundamental distinction between material that introduces or changes the inventory of categories expressed in morphology on the one hand, and the introduction of material that comes in place of existing material without any functional change on the other hand. The first type will be called additive borrowing, the second substitutive borrowing (cf. Breu 1996; Gardani 2008: 22ff.). My impression from a scrutiny of adduced examples of borrowing, excluding the occasional unproductive introduction of foreign morphemes otherwise only found in 2

I know of only one similar case, the use of Greek endings attached to the borrowed diminutive suffix in Cypriot Arabic, which is productive both with nouns of Greek and of Arabic origin (Borg 1985: 125). ˙ Note however that Cypriot Arabic, unlike Sebjan-Küöl Even, has a fully developed parallel system with nouns.

contact-induced change

257

loanwords, is that additive borrowing is much more common than substitutive borrowing (contra Weinreich 1953: 33; Gardani 2008: 84). In fact, this stands to reason: there is no clear functional explanation for the transfer of an isolated morpheme to express something that is already expressed. However, the bilingual speaker confronted with different categorizations in the two languages (s)he uses, may wish to express the same categories in the two languages. This is what has been called ‘feeling the need’ in a seminal article by Eithne Carlin (2006). For example, speakers of Mawayana (Arawakan), living in a Waiwai (Cariban) and Trio (Cariban) environment in Surinam have borrowed a first plural exclusive pronoun from Waiwai. Originally, Mawayana had no inclusive/exclusive distinction. The reason for introducing it was undoubtedly that, being fluent in languages which exhibit this distinction, the speakers of Mawayana ‘felt the need’ to make this distinction also in their language. This does not mean that substitutive borrowing is also rare in other parts of the grammar. Thus, metatypy, as described by Ross (1996) typically involves, among other features, a convergence in syntax which does not necessarily entail changes in categorization. It may very well be that the tendency towards additive borrowing is typical for the introduction of isolated inflectional elements, and that elsewhere in the grammar additive and substitutive borrowing are more equal. The difference between additive and substitutive borrowing is quite clear where it comes to the borrowing of forms. It also applies, however, to the borrowing of patterns. In pattern borrowing, there are cases where a new category (or a changed category) appears in inflection, using native forms, which one could consider additive borrowing. On the other hand, one sometimes witnesses the reshuffling of inflectional elements according to the norms of the contact language without changes in categorization. This could be considered parallel to substitutive borrowing. In the study of inflectional borrowing, emphasis is often laid on the formal features of the borrowed elements (of course, thereby focusing on borrowed forms rather than patterns, and disregarding concomitant borrowing). A number of universal tendencies have been proposed in this connection. The summary outlined in Table 11.5 from Seifart (2012) shows the properties of individual morphemes used to predict borrowability, based on Weinreich (1953), Heath (1978) and Wilkins (1996) (cf. also Gardani 2008). Table 11.5 Properties that predict borrowability Formal properties

Syllabic forms > subsyllabic elements Clear formal boundaries > fusional affixes Structural and syntagmatic unintegratedness > integratedness

Semantic properties

Simple, clear meaning > complex, opaque meaning Non-portmanteau morphemes > portmanteau morphemes Affective meaning > non-affective meaning

Source: Seifart (2012: 474).

258

maarten kossmann

It remains to be studied to what extent these proposals, many of which have so far only been substantiated by limited data, remain useful once the data on which they are based have been subjected to a functional analysis as just sketched above. Seifart (2012), takes a different approach to the question, basing himself on the Principle of Morphosyntactic Subsystem Integrity, ‘which predicts that in situations where various grammatical morphemes are borrowed, these tend to be morphosyntactically interrelated, rather than being random collections of forms or sets of forms’ (Seifart 2012: 471). Thus he observes that morphological borrowings in Resígaro (Arawakan, Brazil) cluster in specific paradigms or morphosyntactic subsystems, while there are only a few borrowings in other parts of the language. Thus, among twenty-eight borrowed suffixes, no less than nineteen are noun class markers—the full Bora set (see also Section 11.6), while most other borrowed suffixes are tightly related to them. He proposes that borrowing the entire set is a way of preserving these subsystems intact. However, one is tempted to consider the clustering of elements in certain subsystems as an entailment of the categorial difference between the subsystems in the two languages in contact. In such a view, the take-over of the noun class markers in Resígaro would be a way to assimilate the Arawakan language (with only little noun class marking) to the Bora structure. The preservation of the integrity of the system would follow from the idea that it is basically the system that is copied (using borrowed materials); in order to do so, one has to copy all elements that make up the system, as far as they were not already there in the language previously.

11.4 Substitutive Borrowing of Inf lectional Forms ............................................................................................................................................................................. Substitution of inflection by borrowing is conspicuously rare, and only few convincing examples can be found. Some cases where substitution of inflection has been proposed concern languages which show etymological asymmetry between lexicon and grammar. In such languages, the bulk of the lexicon comes from one source, and most grammatical features from another source. Assuming that the lexical source constitutes the native part, one could say that in these languages the entire morphology has been borrowed, replacing whatever native inflectional elements there were before. However, it is as easy (and in most cases historically more probable) to say that in these languages a foreign lexicon was introduced into a pre-existing (native) grammatical format, a process referred to as relexification. Most so-called ‘mixed languages’ are of this type; one may mention as examples Ma’á (Cushitic lexicon in a Pare frame; Mous 2003), Media Lengua (Spanish lexicon in a Quechua frame; Muysken 1994) and Old Helsinki Slang (Swedish lexicon in a Finnish frame; Jarva 2008). In such cases, relexification seems to have been a conscious operation by the speakers (Mous 2003), and may have happened in the course of

contact-induced change

259

Table 11.6 Different dialectal variants in Megleno-Romanian verb inflection (present tense, first verb class, ending in certain consonant clusters)

1sg 2sg

I

II

III

Macedonian

antr-u antr-i

antr-um antr-iš

antr-˘am antr-iš

-am -š

‘enter’a

a The translations are not given in Capidan (1925)—he does provide derivations from Latin,

though—and I am therefore not entirely sure about them. Sources: Capidan (1925: 159); Friedman (2002).

a single generation. I shall follow the relexification point of view and therefore not consider the inflectional elements in languages of this type as borrowed (cf. the discussion in Boretzky and Igla 1994: 60–3). Otherwise, substitution of inflection is rare, and most adduced examples are problematic. A classical case of inflectional borrowing (cited in extenso in Weinreich 1953: 32) is found in Megleno-Romanian, a language closely related to Romanian, spoken in a number of villages on the Greek-Macedonian border. In some MeglenoRomanian varieties, the first and second person singular forms of the present tense of one verb class have received Macedonian endings. This only happens in verbs whose stems end in specific consonant clusters. According to Capidan (1925: 159), three different forms appear in the dialects (Table 11.6). The forms under I are regular Romanian forms, and the endings under III are identical to Macedonian inflection. The forms under II present a blend of the two systems: the Romanian ending -u (1sg) is combined with the final consonant of the Macedonian form. The take-over of these forms may have been facilitated by the presence of endings with final -m (1sg) and -š (2sg) in originally Romance forms belonging to other verb conjugations. Final -š in the second person singular is found in a different conjugation class (e.g. conjugation class IVa according to Capidan 1925, dialect type I). (1)

1sg 2sg

sirb-es sirb-eš

(= kick. Verb4: = verb. Verb4: = t a ŋ. Verb4: = a n d a ŋ-.

The treatment of Tübatulabal we have outlined here relies on a separation of the definitions of basic and reduplicated stems on the one hand from the morphosyntax with which they are typically associated on the other. There are a number of different ways one can interpret deponency. Stump (2007) provides a useful typology that distinguishes between form deponency and property deponency. The former is where a word uses the wrong exponence for the morphosyntax associated with it, whereas the latter involves the wrong morphosyntax for its meaning. Müller (2013) adds to this a further two distinctions to the typology of deponency. These are spurious morphosyntactic deponency and spurious morphomic deponency. Both of these involve the claim that deponency is epiphenomenal, because there is no mismatch. In the first case this is because the morphosyntactic feature is more abstract and therefore more flexible in its interpretation. In the second case the exponence is claimed to reflect a morphomic property set, and what is interpreted to be deponency arises because it reflects this underlying morphological systematicity. For our analysis of Tübatulabal there are two different operations for realizing verb stems, and there are two different classes of verbs which happen to contrast in the operations which they access to realize the relevant morphosyntax. The asymmetry of deponency is achieved by overriding the realization in the class lower in the hierarchy, as in (20). We observe this as deponency, partly because there are two opposing feature values involved, and because we selected from two different operations. This comes close to Müller’s notion of spurious morphomic deponency, as the morphosyntax links to purely formal classes defined for the reduplicated and basic stems. In contrast, a purer kind of form deponency would involve at least the following: (25) a. Association of the forms in question with paradigms which extend morphosyntactic feature value x. b. Direct referral of x’s paradigmatically opposed feature value y to x. Even with a small set of mechanisms available for us to model inflectional structure, we are able to make further differentiation within typologies such as Müller’s extension of Stump. We now turn to another phenomenon for which referral is one of the options, namely syncretism.

modelling inflectional structure

295

12.4.4 Syncretism In Section 12.3.2 in our discussion of Polish we have already seen a treatment of syncretism. By default, for instance, the dative singular is treated as being the same as the locative singular, as in (26). (26)

MOR_NOUN: == MOR_NOMINAL == "" ...

In (26) we have a referral of the dative singular to the locative singular. Referrals are discussed in Stump (2001b: 212–41) and Brown and Hippisley (2012: 167–80). A well-known alternative treatment of syncretism is to use underspecification so that the features are broken up further. In this case locative and dative are broken up further and a feature is used to pick out an identifiable Gesamtbedeutung in the sense of Jakobson (1936[1971]/1958). As we have seen in Section 12.2, the additional distinctions made by these features are irrelevant for syntax. In contrast, in (26) the need for additional features is obviated by the use of a particular rule type, the referral. The justification for this is that there can be an asymmetry in that the source of the referral is considered in some sense to be the basic meaning of the exponent (see discussion in Baerman et al. 2005: 136–50). There may, however, still be a need for underspecification involving features which are relevant for syntax. For instance, because gender is neutralized in the plural in languages like Russian, we can treat the whole gender feature as underspecified in the presence of plural number, and as this implies a particular structure to the paradigm, we can use this to account for some of the gender syncretism in the singular in Russian adjectives (between neuter and masculine). Finally, we will still have to resort to the use of morphomic indexes in dealing with syncretism where there is no natural class that can be identified and no indication that the patterns of syncretism are directional (Brown and Hippisley 2012: 180–3). Syncretism may therefore involve referral, underspecification, morphomic indexing, or a combination of these, and computational modelling helps us see those mechanisms it shares with other phenomena, such as deponency.

12.5 Conclusion

.............................................................................................................................................................................

Computational modelling brings with it a number of advantages. We can check to see that our analyses make the correct predictions. Connected with this we can verify that the theoretical adjustments made for a particularly demanding aspect of the morphology do not have undesirable consequences for other parts of the inflectional system. Subtle distinctions with wide ranging consequences can be discovered, as well as a better understanding of the basic connections between related phenomena.

chapter 13 ........................................................................................................

MACHINE LEARNING OF INFLECTION ........................................................................................................

katya pertsova

13.1 Introduction

.............................................................................................................................................................................

The task of learning a natural language is an instance of inductive inference (making generalizations based on past observations to predict future ones) studied within several traditions in the sciences under such names as ‘machine learning’, ‘grammar induction’, and ‘computational learning theory’. In this chapter, I aim to introduce some of the concepts and approaches used in the formal study of learnability and the relevance of this field to theoretical linguistics. I then discuss issues specific to learning inflectional morphology. However, my focus is not on providing a comprehensive review of the proposals for learning inflection, but rather on the general problems and obstacles in learning inflection and on the question of what properties a computational system needs if it is to overcome these obstacles. Inflection, considered separately from other components of language, is relatively restricted in its expressive power which, on the one hand, makes it easier to learn than syntax. On the other hand, inflectional systems are full of irregularities and mismatches between different levels of structure, and such irregularities make learning difficult. Overall, I conclude that linguistically interesting proposals for machine learning of inflection should provide explanations for the nature and extent of irregularities present in inflection and make reasonable predictions for the specific patterns of language acquisition and language change. That is, in addition to absolute universals (if there are any), formal learning models can further our understanding of [soft universals or tendencies] which might arise as a result of learning biases. A prominent leitmotif of this chapter is a focus on how computational learning models can serve as a valuable tool for constraining linguistic theories and for uncovering deeper explanations for a variety of linguistic phenomena. The relevance of

298

katya pertsova

formal learning models to linguistic theory stems from the causal connection between learning and language structure standardly assumed within the rationalist approach to language (Chomsky 1975). On this approach, the properties we find in human languages at least to some extent reflect constraints on our cognitive capacity including constraints on the language-acquisition mechanism.

13.2 Goals of Computational Learning Theory ............................................................................................................................................................................. In the machine learning literature, a language learner is a function that maps data to hypotheses about the data source (e.g. grammars) or to probability distributions over a set of such hypotheses. This learning function can be defined by an algorithm of any sort: a neural network, a stochastic process, a Boolean classifier, an inductive logic program, and so on. The field of computational learning theory as developed by Gold (1967); Blum and Blum (1975); Angluin (1980); Valiant (1984); Jain et al. (1999) and others is the study of such functions and their behaviours within frameworks that explicitly specify both the criteria for the evaluation of learning success and various constraints on the learning situation, such as what kind of input is allowed, how it is presented, and what the resource limitations are. As in other sciences, idealizations are deliberately used to focus attention on specific aspects of the problem abstracting away from other details. Within more data-oriented traditions including much of the statistical learning literature, the research goals are somewhat different—typically, to directly model a certain aspect of human behaviour or corpus distribution. However, the same mathematical results apply to these models and the models themselves can be easily translated into the computational learning theory terms (Niyogi 2006). Learning is understood as choosing the right or the most probable hypothesis about the data from a set of possible hypotheses (where the right hypothesis is the one that generates the target language with probability 1). The learning problem is sometimes viewed as the problem of searching for the best grammar within a space of possible grammars from some limited class. Learning algorithms are said to learn a class of languages if they can identify every language in the class given some data from this language. In terms of first language acquisition, learning a class of languages (rather than a single language) corresponds to the idea that human learners must be able to learn whatever natural language they happen to be born into. For a non-technical linguist-oriented introduction to some key concepts and controversies within the field of learnability I refer the reader to Heinz (2013). In the next sections I highlight one of the central results of learning theory—the fact that no learning algorithm can learn an unrestricted class of languages. The restrictions on the class of languages (or on the class of grammars) to be learned constitute knowledge that is present prior to learning. These restrictions are often referred to as learning

machine learning of inflection

299

biases (which can also be viewed as constraints of Universal Grammar (UG)). The fact that learning without any prior knowledge is impossible, should put to rest the debate between the nativists and the radical empiricists, although the related debate between the domain-general vs. domain-specific constraints on language acquisition remains very vibrant (Pearl and Lidz 2009; Clark and Lappin 2011; Perfors et al. 2011; and others).

13.2.1 Learning without Restrictions is Impossible Language learnability provides an obvious constraint on what constitutes a possible human language—such a language must be learnable. One natural starting point for investigating which classes of formal patterns are learnable is the Chomsky Hierarchy (Chomsky 1956), which defines a set of increasingly complex classes of rewrite grammars corresponding to classes of languages generated by these grammars: finite languages, regular languages, context-free languages, context-sensitive, and recursively enumerable (or r.e.) languages.1 The importance of the Chomsky Hierarchy is that it establishes a certain metric of complexity for languages in terms of the complexity of the mechanisms for their recognition and generation, which one might also expect to correlate with the learning difficulty. The current consensus is that at least some human languages are mildly-context sensitive in the sense of Joshi (1985), a class of languages intermediate in complexity between context-free and context-sensitive languages (Joshi 1985)2 or, perhaps, even more complex as recently suggested based on certain kinds of copying dependencies in syntax (Stabler 2004; Kobele 2006). However, one of the early controversial results in learnability theory due to Gold (1967) is that even relatively simple classes within the Chomsky Hierarchy, such as the class of regular languages, are not feasibly learnable (that is, learnable in a realistic amount of time) in many frameworks. For instance, regular languages are not learnable in Gold’s framework ‘identification in the limit from full positive texts’. This framework allows the learner an unlimited amount of time and computational resources, but requires exact convergence (that is, learning is defined as reaching a point after which no errors are ever made). Subsequent work made clear that it is not the stringent exact convergence requirement that makes learning large classes of languages difficult. Modifying the above framework so that success is evaluated in a probabilistic fashion does not change any of its negative results (Pitt 1985; Wiehagen et al. 1984). Similarly, none of the classes in the Chomsky Hierarchy are learnable in a probabilistic framework known as Probably Approximately Correct learning (PAC) which is rooted in the principles of statistical learning with older history (Valiant 1984; Kearns and Vazirani 1994). 1 For more details on the formal language theory and the Chomsky Hierarchy the reader may consult Partee et al. (1990); Hopcroft et al. (2001). 2 More specifically, these are context-sensitive languages that are efficiently parsable, have limited crossing dependencies, and have constant growth (the growth of string length is linear).

300

katya pertsova

The negative results mentioned above do not indicate that the mathematical study of learnability is useless and not revealing, as is sometimes suggested. These results should be considered together with a large number of positive results showing that specific classes of languages with well-defined structure are learnable in various frameworks. Moreover, the whole class of r.e. languages becomes learnable (but not efficiently so) in certain settings that either enrich the learner’s input or restrict the types of data on which the learner is required to succeed (Gold 1967; Horning 1969). What these results tell us is that while learning unrestricted classes of languages is difficult or impossible, restricting assumptions about the nature of the input or about the hypotheses available to the learner dramatically improve the learning outcome. Non-trivial (i.e. generalizing) learners have been proposed for patterns similar to those found in human language (Yokomori 2003; Clark 2010; Heinz 2010). These learners are defined for restricted classes of languages that do not correspond to any specific class in the Chomsky Hierarchy, supporting the growing consensus that the class of human languages (i.e. languages that can be learned by humans) corresponds to some cross-section of the Chomsky Hierarchy. The exact nature of restrictions on this class is still a matter of discovery—research that will surely benefit from a joint effort on the part of linguists and the computational learning community. What the latter community contributes to this research are precise formulations about what guarantees learnability in specific settings and insights about what structural properties of the learning problem drive generalization. What linguists and psychologists can bring to the table are the empirical facts that help better identify structural properties forming the core of human languages (properties that learners can exploit), as well as facts that help make learning proposals more psychologically plausible.

13.2.2 What do Computational Learning Models Tell us About Language Many of the computational learning models should not be taken as modelling exactly what is going on in the brain (although some models, such as neural networks, are claiming to do exactly that). I would like to stress that this fact does not make such models uninteresting or irrelevant for linguists and psycholinguists. Computational learning models help us gain insight about what makes learning possible and answer a variety of important questions starting with really general ones, such as ‘what classes of languages are in principle learnable by any learning algorithm’ to more specific ones such as ‘what effect a particular bias (say, an expectation that no two words can have the same meaning) has on learning’. Additionally, restrictive learning models can help us explain and hone in on the set of properties that languages share in common (both categorical and ‘soft’ statistical universals) by highlighting which properties of grammars or languages aid in the learning process. On the other hand, some very powerful unrestricted learning models

machine learning of inflection

301

capable of learning the entire class of r.e. languages, are not very insightful in the sense that they do not explain why the class of natural languages is vastly smaller than the class of r.e. languages. Thus, the traditional connection between linguistic universals and learning biases plays a central role in model design and selection—we assume that some of the restrictions in the data are there because they make learning easier. For example, in statistical learning literature, the simplicity bias (discussed further in section 13.5.1) falls out from the assumption that simpler hypotheses are preferred to more complex ones. In these models ‘restrictiveness’ is discussed with respect to attempts to achieve a balance between simplicity and goodness of fit (the ability to account for the data as tightly as possible). The more restricted and simple the language class, the more likely it is that a simple hypothesis capturing the data also fits the data well. On the other hand, less restricted, complex classes are hard to capture with simple models without running the error of underfitting (or overgeneralizing, which is one way of giving up global goodness of fit); and vice versa: choosing a model that fits well all the sample data leads to proposing highly complex hypotheses that run a risk of overfitting (or undergeneralizing, which is another way of giving up global goodness of fit). Thus, the best scenario for a statistical model is the one in which the language class to be learned is sufficiently restricted along the lines that accord with the model’s notion of simplicity, and, as a result, are describable with high accuracy by a simple grammar. In other learning models, simplicity biases and other linguistic restrictions can be directly built into the assumptions about the hypothesis space or the learning algorithm. The rest of this section briefly addresses a common concern that the assumed relationship between language universals and learning biases is faulty. There are different ways in which the connection between learning biases and typological preferences can be fleshed out. One possibility (in rough terms) is that in the process of language change, which often involves the co-existence of several alternatives (Weinreich et al. 1968), the alternatives for which the learner is biased are learned more faithfully and by a larger number of new learners because they require less effort. As a result these alternatives are more stable and eventually drive out the other competitors. This does not lead to complete obliteration or simplification of all irregularities, as new complications are introduced all the time by processies which may be ‘blind’ to certain components of grammar. One might object, however, that linguistic universals (whether categorical or statistical) might be due to factors not related to learning, such as historical accidents, physiological, and functional constraints. There are several responses to this objection. The commonly suggested alternatives to learning biases in the domains of morphology and syntax (e.g. communicative and processing constraints) are only alternatives to domain-specific biases, but not necessarily to domain-general biases. Many communicative pressures can be restated as learning pressures because what hinders communication also hinders learning which relies on communication for transmission of data to the learner. Anderson (2008) also observes that it is possible for some preferences that had their origins in a functional explanation to become encoded as

302

katya pertsova

innate preferences in the architecture of the learning mechanism through the so-called evolutionary ‘Baldwin effect’, the propensity of a learned trait to become innate over a long period of time. Finally, on the empirical side, experiments on artificial grammar learning and on evolutionary simulations of language change support the role of learning biases in emergence of linguistic universals including biases grounded in communicative factors (see Kirby et al. 2014 for a recent overview of experimental and computational methods used to study iterated learning over time, and how it shapes linguistic structure). Thus it is highly plausible that some (and perhaps many) language universals are due to learning biases. Therefore, an insightful and explanatory learning model will be the one that is appropriately restrictive by incorporating such biases. To my mind, the difficulty of securely establishing the source of linguistic universals only underscores the importance of formal learning models, especially when they are embedded in broader computational models of language change. Such models can help us quantitatively test the load of different factors in explaining language universals since different factors in the model (e.g. the learning mechanism vs. the population dynamics) may contribute differently to a number of learning-related predictions, such as predictions for errors during language acquisition, rates, and directions of language change, side effects on other aspects of language structure, language shifts in bilingual populations, and so on. See Niyogi (2006) for an exploration of this topic in the context of an evolutionary model that includes factors of communicative efficiency, population dynamics, and learning algorithms inspired by linguistic proposals. To sum up, one of the central imports of modelling language learning is similar to the major goal of linguistic research, that is, the goal of identifying restrictions on linguistic patterns and providing explanations for the source of linguistic universals.

13.2.3 Finer-grained Predictions of Learning Models Goodness of fit is the main criteria for evaluating learning models. But for models that strive for psychological plausibility there are additional criteria that have to do with how closely the model encodes the learning situation of a human child and fits children’s learning trajectory. This includes assumptions about what computational resources the model relies on, whether it has access to negative evidence, whether it can handle noise, what kinds of errors it makes, and so on. To the extent that such assumptions lead to concrete predictions about the speed of learning and the error patterns, they can be tested against acquisition data, which provides finer-grained evidence for learning models than mere facts about grammaticality. Below I review just a few examples of the kinds of considerations that arise in evaluating models with respect to their psychological plausibility. The first consideration concerns memory resources. Most learning proposals implement some version of memory, a repository of the data available to a learner. Some models have access to every single data point in addition to computing compressed

machine learning of inflection

303

representations of the data, some models are completely memoryless, and others are in-between these two extremes allowing for various levels of memory. Assumptions about memory resources and how they are used have consequences for the predicted pattern of errors during language acquisition. These predictions along with the psycholinguistic findings about how memory is used in language learning can be used to decide on the most realistic type of model. For example, several psycholinguistic studies based on frequency effects suggest that even regular words of relatively low frequency are stored and accessed by language users (Alegre and Gordon 1999; Baayen et al. 2003). Such studies are taken to provide support for the memory-based models. However, memory-based models place a heavy burden on storage and in this respect they are less ideal than memoryless learners. A memoryless learner is an incremental function that at each step maps a current data point and a previous hypothesis to a new hypothesis. Several syntactic triggering learners (inspired by the Principles and Parameters framework) are of this type—after seeing a new data point they may change some parameter setting which then produces a new grammar (Gibson and Wexler 1994; Yang 2002). These models provide greater storage economy which is particularly important for learning infinite or very large classes of patterns. The absence of memory, however, makes learning many problems significantly more difficult. For instance, Fodor and Sakas (2005) discuss how memoryless strategies have certain difficulties with learning languages that are in a subset relationship. Certain kinds of memoryless learners also have a problem with local maxima (discussed in Section 13.2.3). In morphology, there are proposals for dual-route learning models that implement storage for the irregular subsets of the data and compute compressed descriptions (e.g. rules) for the rest of the data (Pinker 1999). Research by Michael Ullman and his colleagues suggests that these two types of representations correspond to two different types of memory humans are known to have: the declarative memory (storage of facts) and procedural memory (storage of procedures, rules) (Ullman 2001a). But experimental data from different labs is not consistent: some support the dual-route model while others do not. Thus, how much memorization and how much generalization humans perform when learning morphology remains an open empirical question. Another important way of evaluating psychological plausibility of a learning algorithm is to look at the general trends for error patterns produced in the process of learning. One such trend is conservatism. Gibson and Wexler (1994) loosely define conservatism as a property learners have if they ‘make no large-scale reorganizations of their grammars in a single learning step’ (Gibson and Wexler 1994). They propose that conservatism is a desirable property in a psychologically plausible learning model because children are thought to learn in a very gradual piecemeal fashion.3

3 It is important to distinguish between two different kinds of conservatism: conservatism with respect to the grammar, and extensional conservatism which is conservatism with respect to the languages generated by the grammar. Children may display large jumps in their language production, suddenly producing more different types of expressions, but as long as these jumps are due to a small change in the grammar, they are still conservative in the first sense.

304

katya pertsova

A widely used conservative learning method is the so-called ‘hill-climbing’ method. This method relies on an incremental memoryless optimization strategy that proceeds by making small adjustments to a current hypothesis towards a more ‘optimal’ solution, namely the one that better explains the current data point (one example is the Gradual Learning Algorithm (Boersma 1997) for learning stochastic Optimality Theory). However, although such learners meet the definition of conservatism above, they face a notorious problem if the hypothesis space contains local maxima. A local maximum is a hypothesis that cannot be improved upon by the hill-climbing algorithm given any possible input, and yet it is not the most optimal global solution to the learning problem. This scenario is often illustrated through an intuitive example of actual hill-climbing. Suppose that to get to the top of a hill a climber only takes steps in the direction that will land him at a higher elevation than his current position. This strategy will guarantee that eventually the climber will reach the top of the hill, but not necessarily the highest hill since once he reaches the top of any hill, he will stop (because a step in any direction will be a step down). Therefore, hill-climbing methods can only be applied if the set of possible hypotheses for the target language class does not contain local maxima. Do grammars of human languages have local maxima? This, of course, depends on what these grammars are like. For instance, the exploration of some parametric hypotheses spaces (in which a grammar corresponds to a set of parameters as in theories of Government and Binding (GB) or Head-Driven Phrase Structure Grammar (HPSG)) reveals that local maxima are problematic (Niyogi and Berwick 1996). Thus, depending on the theory of grammar we adopt, a conservative hill-climbing method may not be the best learning strategy.4 There are other facts about human language acquisition we may consider to be important for learning models to replicate. For instance, we may require that a learner be consistent. A consistent learner is a learner whose hypotheses account for all the data seen so far. A consistent learner is not necessarily conservative in the sense above, but it guarantees that the hypotheses it produces at each step agree on all the received data. There are good reasons to think that human learners are generally consistent, that is, they do not gratuitously dismiss the data that they have seen in the past. Another example of a learning property associated with a particular error pattern is monotonicity. A monotonic learner is a learner whose current hypothesis is always a superset of the previous hypotheses. That is, a monotonic learner is a learner that never reverses its predictions. Such learners are mathematically nice because they make it easy to reason and to prove facts about a learner’s behaviour. However, it is likely that human learners are not monotonic because children occasionally overgeneralize and later correct their overgeneralization errors (Ervin 1964; Marcus et al. 1992). This implies the existence of a mechanism that allows the learner to retreat from its previous hypothesis, producing a non-monotonic pattern of generalizations. The linguistic descriptions

4 However, there are various methods for augmenting hill-climbing with additional assumptions (e.g. memory, restarts, stochastic modifications) which alleviate the local maxima problem.

machine learning of inflection

305

involving ‘default’ and ‘elsewhere’ rules are examples of non-monotonic reasoning systems that could be thought of as arising from a non-monotonic learning strategy. If human learners are indeed non-monotonic, this would already rule out many learning algorithms from our consideration. To conclude this section, computational modelling of language learning is also useful for exploring the effects that the learning environment and the limitation on computing power may have on the learning outcome.

13.3 Inf lection as a Formal System

.............................................................................................................................................................................

Turning from general issues in machine learning, we now consider the learning of inflectional morphology in particular. What kinds of learning techniques and learning biases are appropriate for inflection? Before these questions can be answered, it is helpful to first clarify what is learned when learning inflection, and how complex this task is.

13.3.1 Expressive Power of Inflection There is no consensus on the exact criteria separating inflection from other types of morphological marking such as derivation and cliticization (Stump 1998), but we may roughly define the inflectional component of a language as a set of strings resulting from taking a union of word forms of every lexeme in the language. Word forms of the same lexeme do not differ in their core semantic meaning (i.e. the concept expressed by the word) and in their lexical category, but differ only in morphosyntactic properties (whatever our current understanding of such properties may be). Additionally, a single word form does not typically contain multiple instances of the same inflectional morpheme.5 To evaluate the complexity of inflectional patterns, we can remain agnostic about whether words are derived in syntax or in a distinctly different morphological component of grammar. Let us now consider this question: what is the simplest grammar that accounts for the set of inflectional forms of a language in an insightful way? And, relatedly, what types of learning algorithms are particularly fitted for learning this grammar? In the simplest scenario, the set of inflectional word forms is finite (supposing finite vocabulary for the moment). The simplest kind of representation for a finite set is an unstructured list, and the simplest way to learn such a list is memorization. However, inflection cannot be plausibly learned by memorization alone. In subsection 13.3.1.2, 5 One possible exception to this claim is the apparent recursion known as ‘Suffixaufnahme’ when the possesive NP carries the case and number inflection of the possessor (Plank 1995).

306

katya pertsova

I consider proposals that inflection can be fully handled by relatively simple computational devices, namely finite-state automata (which generate regular languages) and which are widely used in computational models of morphology (Sproat 1992; Beesley and Karttunen 2003; and others). The upshot is that simple learners that can handle subsets of regular patterns are sufficient for handling inflection.

13.3.1.1 Against lists and memorization Since the number of inflectional exponents in any language is finite, and since inflection is not recursive, one could in principle list all the word forms in a language (assuming, for the sake of an argument, that speakers know only a finite number of words). Such a list could be derived in the process of simple memorization. However, most morphologists would agree that memorization alone (without generalization) is unrealistic as a model for how humans learn inflection. There are several reasons commonly cited for thinking this is the case. First, we know that speakers generalize as is evinced by children’s errors, wug-tests, and loanword adaptations. Additionally, we know that sets of word forms, though finite, can be extremely large in richly inflected, especially agglutinative, languages (Hankamer 1989; Kurimo et al. 2006). Moreover, the distribution of data within the speakers’ experience is sparse. That is, speakers mostly experience the same patterns repeated over and over again, while majority of patterns are rare. This is known as the Zipfian distribution, which was originally established for word frequencies (Zipf 1932), but has since been found to hold for many other linguistic phenomena including inflectional morphemes (Chan 2008). For example, as Table 13.1 based on (Chan 2008) shows, many word forms in richly inflected languages are exceedingly rare as can be seen from the fact that available corpus data do not include the full verbal paradigm for any lemma. The last column, percent saturation, shows the maximum percentage of verb forms attested in a corpus for a single verb lemma. For all languages in the sample, except English (which has only six maximally distinct verb forms), saturation is below 100 per cent; and for some languages it is only around 50 per cent. Zipfian distribution renders the memorizing learner implausible: if most expressions are rare, one has to wait a very long time until all of them are heard. There are other, perhaps more interesting, asymmetries in the data making certain types of patterns more frequent than others. Just to mention a few examples: free-variation of morphemes appears to be rare (Kroch 1994); syncretism is more common in the marked portions of a paradigm (Haspelmath 2002); when allomorphy is conditioned by other material in the word, this material will almost always be closer to the root (Carstairs-McCarthy 1987). Supposing that some of these facts are the results of learning biases, a learner encoding these biases will converge on the right grammar much earlier than a memorizing learner. For instance, a learning algorithm that is biased against synonymy or free variation, will be more efficient than a learner without such a bias since, when deciding on a meaning of a new morpheme, it will not have to entertain meanings already assigned to

machine learning of inflection

307

Table 13.1 Sparseness of paradigms in corpora (from Chan, p.79) Millions of tokens

No. of total verb forms in corpus

Max. no. of forms for any lemma

% saturation

1.2 1.3 0.6 2.0 2.1 2.8 1.2 2.5 2.4 2.6 1.0 1.7 1.4 1.4 0.3 0.3

6 6 22 72 365 83 76 33 32 51 21 45 55 55 39 49

6 6 16 41 147 45 48 23 24 34 14 33 47 46 27 31

100.0 100.0 72.2 56.9 40.3 53.2 63.2 69.7 75.0 66.7 66.7 73.3 85.5 83.6 69.2 63.3

Corpus English(Brown) English(WSJ) Basque Czech Finnish Greek Hungarian Hebrew Slovene Spanish Swedish Catalan Italian CHILDES Spanish CHILDES Catalan CHILDES Italian Source: Chan (2008: 79).

other morphemes (see a simulation in Taguchi et al. 2006). Generally, whenever a finite set is structured (even if only in a probabilistic sense so that some patterns are more frequent than others), this gives a generalizing learner an advantage over a memorizing learner which can just as easily memorize unstructured data as it does structured data.

13.3.1.2 The upper bound on inflectional complexity As mentioned in Section 13.2.1, the set of human languages is currently believed to be some subset of mildly context-sensitive languages. However, some aspects of human languages are certainly more restricted than that. Most relevantly, there are reasons to think that inflectional patterns are simpler than phrasal syntax. First, inflection does not allow recursion. As Haspelmath (2002) points out, nothing logically rules out recursion in inflection as one can imagine hypothetical forms like cat-s-es meaning ‘sets of cats’. But such iterated inflections are virtually unattested. Second, unlike syntactic dependencies, most word-internal inflectional dependencies tend to be local, that is, conditioned by linearly adjacent material or by structurally adjacent material with limited amount of non-local dependencies (more on this in Section 13.5.2). Third, the order of inflectional morphemes within a word is often fixed—many languages have the so-called templatic structure where the same morphosyntactic features are realized in the same order. Fourth, languages do not show true scrambling of

308

katya pertsova

inflectional affixes, although occasional restricted patterns of variation in affix order are possible. Fifth, the semantics of inflection is relatively primitive, corresponding mostly to taking (multi)unions of inflectional features in contrast to a variety of complex semantic functions at the sentence level. And finally, at the level of the mapping between form and meaning, inflectional elements realized in the same position typically appear in complementary distribution: for instance, a nominative case marker cannot appear where dative is expected and vice versa. On the other hand, ‘open-class’ lexical items are less restricted in their distribution: languages are full of words with overlapping meanings at various levels of generality. The speakers can choose among these levels at will (e.g. in exactly the same situation one can choose to refer to an instance of a rose as ‘a rose’, ‘a flower’, ‘a plant’, or ‘an object’). It is because inflection, and morphology more generally, is more restricted than phrasal syntax, that relatively simple computational models of morphological grammars and morphological learning, such as finite-state models, have been so successful (Koskenniemi 1983; Theton and Cloete 1997; Beesley and Karttunen 2003). These models represent the data as a finite-state automaton, a mathematical abstraction used to generate or recognize sets of expressions that require only finite memory and that correspond to the regular languages in the Chomsky Hierarchy. As exemplified in Figure 13.1, an finite-state automaton is standardly drawn as a directed graph in which memory states (represented by circles) are connected to each other by labelled arrows. One can traverse an automaton graph by following the arrows from an initial to a final state, outputting or reading the labels on arrows along the way. Figure 13.1 is an example of a finite state transducer (an automaton in which symbols correspond to input-output pairs) for a small set of Russian word forms, the nominative possesive pronouns.

moj : ‘mine’ tvoj : ‘yours’ ø : masc

nash : ‘ours’ 2 vash : ‘y’all’s’ 1

-a : fem

chej : ‘whose’ ix : ‘theirs’ jevo : ‘his/its’

-o : neut 3

jejo : ‘hers’

fig. 13.1 A finite-state-transducer for Russian possesive pronouns. The initial state is an ellipse, the final state—a double circle. This FST captures the fact that 3rd person possesive pronouns behave differently from other pronouns by not taking gender agreement suffixes.

machine learning of inflection A B C … A

B

309

C …

fig. 13.2 Crossing dependencies in reduplication

Finite-state models can generate or recognize infinite languages (since they allow loops), but only of limited complexity—namely, those languages in which the identity of each symbol depends on finitely many preceding symbols. It is because inflectional patterns do not require infinite or arbitrary amounts of memory (with a possible exception of reduplication) that they can be described using finite-state automata. This contrasts with certain syntactic patterns (e.g. centre-embedding) that require an arbitrary amount of memory. Finite-state methods are not limited to concatenative morphology. Computational linguists, relying on proposals of theoretical linguists such as McCarthy (1981), have developed finite-state models that can handle non-linear templatic morphology of Semitic languages. For instance, Kiraz (2000), building on a proposal by Kay (1987), presents a general finite-state framework for computing templatic patterns using multi-tape finite automata implementing the idea of multiple tiers in autosegmental phonology. Thus, morphology of Semitic languages can be learned if a learning model is built to expect templatic patterns and has some way of determining when a language is using a concatenative or a templatic method of morpheme combination. One possible challenge for finite-state models of morphology is reduplication since it is best implemented as a copying operation. Unbounded copying, however, is beyond the power of finite-state/regular grammars because it requires keeping track of a potentially unlimited number of crossing dependencies (as shown Figure 13.2). Whole-word reduplication is particularly problematic because the model has to work for arbitrarily long words. (Of course, we can build an automaton that allows reduplication of words up to some sufficiently large number of segments n, but such an automaton would be extremely complex and redundant.) Interestingly, whole-word reduplication intuitively seems to be the least problematic type of reduplication for human learners since the constraint on the size of the reduplicant is very simple. This suggests that either representations of higher complexity (i.e. those that can describe context-sensitive patterns) are involved in learning some aspects of morphology, or that reduplication is best handled not by copying of phonological segments, but by doubling of a single substring within a word. To sum up, while we can in principle account for inflection by listing all inflectional forms, the more elegant, succinct grammars of inflection capturing regularities in the data are more complex, but likely no more complex than regular (or finite-state)

310

katya pertsova

grammars. Since finite-state languages are relatively simple and well-studied, there are a number of efficient and robust methods for learning subclasses of such languages (for examples of learning proposals in morphology that rely on finite-state representations see Kiraz 2000; Clark 2001; and many others).

13.3.1.3 Modularity within language Lurking behind the discussion in the previous subsection is the issue of modularity. The issue is whether the human acquisition device is ‘modularized’ so that more restricted learning procedures are used to learn more restricted aspects of grammar. If inflection can be described with less expressive machinery than syntax, could it be that it is also learned by simpler, more restricted learners? Remember that we have not committed ourselves to the view that inflection in fact corresponds to a separate grammar module. It could very well be that it does not, in which case we can assume that the same learning algorithm that is used to learn more complex syntactic patterns is also used to learn inflection (or both inflection and derivation). This possibility is parsimonious in the sense that it reduces the number of different learning strategies one has to assume. On the other hand, modularity can be argued to be beneficial for learners: namely, it could be more advantageous to have different learning algorithms for aspects of grammar that differ in complexity, just as it is advantageous for a golfer to carry different golf clubs, each tailored for a specific situation. An advantage of a modularized learner over an all-purpose learner is that such a learner can learn certain simpler aspects of grammar before the more complex ones, something that can simplify learning at higher levels of structure. This advantage exists because learning is more efficient within more constrained hypothesis spaces as was discussed in the first half of this chapter.

13.4 Challenges for Inf lectional Learning Models: Ambiguity and Irregularity ............................................................................................................................................................................. Although I suggested that learning the set of inflectional word forms of a language may be relatively simple, it is still far from trivial. One of the main obstacles for this task is ambiguity. Ambiguity involves a non one-to-one mapping or a mismatch between the two components. Inflectional mismatches exist at various levels of representation: at the level of the mapping between phonological form and grammatical function (e.g. null morphemes, syncretism, allomorphy), at the level of mapping between word-positions and features typically realized in these positions (e.g. portmanteau morphemes, variable affix ordering), at the level of mapping between morphological form and syntactic/semantic function (e.g. deponency, heteroclisis), and at the level of mapping morphological constituents to syntactic or semantic constituents (the so-called ‘bracketing paradoxes’).

machine learning of inflection

311

Why are mismatches problematic for learning? In short, because they enlarge the learning space (the possible hypotheses the learner has to entertain) and because in some cases these hypotheses become hard to tell apart from each other. At one extreme, mismatches may make learning extremely difficult or impossible. For instance, imagine learning morpheme meanings in a language in which every morpheme is homonymous or has a million synonyms. On the other hand, when there is no ambiguity at all, learning something about one component of the grammar automatically tells you something about the other ‘matched’ component, and this fact can be exploited by relatively simple generalization techniques. In other words, one-to-one mappings allow the learner to ‘bootstrap’ when learning interlocking components. Perhaps the most common and widely discussed inflectional mismatch is allomorphy, the phenomenon in which the same morphosemantic features are realized differently. Particularly problematic is inflectional class allomorphy in which inflectional classes are lexically specified: that is, membership in a class is not fully predictable from any of the members’ properties. The different realizations of the past tense in English is a good example. The learning of English past tenses is probably the most studied topic in the machine learning of inflection, a topic that sparked the debate between the connectionist vs. symbolic approaches to learning, as well as the debate between the dual route (symbolic rules for regulars + associative memorization for irregulars) vs. single route (associative learning or rule learning for both regulars and irregulars) mechanisms for learning and storing morphological patterns (Ling and Marinov 1993; Rumelhart and McClelland 1996; Pinker and Prince 1988; and others). A prominent feature of lexical allomorphy is that, while it is not fully predictable, many allomorph classes display patterns of semi-regularity. That is, the lexical items that select the same allomorphs may have a Wittgensteinian ‘family resemblance’ structure with overlapping and crossing patterns of similarity without sharing a single set of necessary and sufficient features. An artificial language learning study reported in Brooks et al. (1993) found that fully arbitrary gender-like inflectional classes were not learned by adults and children, unlike classes in which a subset of nouns shared a set of phonological features. These facts suggest that there are statistical restrictions on the amount and arbitrariness of lexical allomorphy. The same can be said about another prominent example of ambiguity—syncretism. Syncretism refers to a situation in which different morphosyntactic features are realized by the same inflectional exponent.6 The simplest examples of syncretism are structured and can be described with feature underspecification, but the more complicated cases exhibit the same semi-regularity as allomorphy: the cells that are syncretic do not share a set of necessary and sufficient properties, yet most of the time they also do not comprise a completely random set. For instance, most patterns of syncretism that 6

Syncretism is often assumed to arise for principled reasons, in the words of Carstairs-McCarthy (1987) ‘in the sense of being more than mere accidental by-products of phonological processes or morphological “spell-out” rules’. Namely, forms that are realized syncretically are also expected to have related function.

312

katya pertsova

cannot be captured by underspecification alone, can be captured by underspecification and blocking (Pertsova 2011). There are also cross-paradigmatic regularities involving syncretism: in many cases, the same cells tend to be syncretic in different paradigms within the same language revealing that something systematic is going on. These two examples of mismatches reveal a general pattern: while ambiguity is pervasive, it is still restricted. In a finite domain with few categorical universals, such restrictions become the defining properties of the system. Therefore, the main theoretical goal of modelling inflectional learning is to predict restrictions on ambiguity (and irregularity) and explain their underlying causes.

13.5 Learning Biases in Acquisition of Inf lection ............................................................................................................................................................................. In this section I discuss some learning biases that are used in many models of inflectional learning. These biases reflect the researchers’ hypotheses about what regularities in the data can be exploited by learners. We will see that some such biases allow learners to successfully generalize in the presence of ambiguity and at the same time explain why this ambiguity exists.

13.5.1 Simplicity The first bias I discuss is based on a general idea of simplicity or economy implementing the principle of Occam’s razor, familiar throughout science. Most learning models, as well as theories of inflection, invoke some version of simplicity or economy constraints in their architecture. In this section, I first introduce how simplicity may be understood and implemented, and then discuss how this notion is used in models for learning inflection. I will also discuss how trade-offs in simplicity at different levels of grammar may introduce the sort of ambiguity we must account for. It is often remarked that simplicity is a subjective and vague concept. Correspondingly, to say that the model is driven by simplicity is not to say much—one needs to specify how this notion is to be understood. Statistical learning models get certain simplicity biases automatically from the kinds of methods they presuppose. For example, models that rely on the Minimum Description Length (MDL) principle (Rissanen 1978) are based on the idea that simplicity can be measured as the size of the hypothesis and the size of the data described with this hypothesis (within a fixed description language). Shorter hypotheses are preferred matching the fact that sets with regularities can be compressed and hence have shorter descriptions. Other statistical models are based on the idea of selecting not the shortest but the most probable hypothesis. However, in many cases this amounts to the same thing since there is a close correspondence

machine learning of inflection

313

between the length of a hypothesis and its probability (for instance, see Vitányi and Li (2000) for conditions under which Bayesianism and MDL converge). While we cannot really know whether these are the right measures of simplicity, the success of statistical models that maximize probability of the data or minimize the hypothesis length on some linguistic tasks, such as morpheme and word segmentation, supports the relevance of the simplicity measures they adopt. In non-statistical models, the simplicity bias has to be explicitly built in. Some researchers prefer building-in simplicity bias ‘from scratch’ as it allows them greater control in interpreting and implementing this bias given the particulars of the data. As one specific example of how a simplicity bias may aid in learning inflection, consider a Bayesian classification model proposed in Frank et al. (2008). The model in question is a general categorization model that the authors apply to the learning of artificial languages in the lab and to the learning of inflectional rules like the English past tense. The rules in this model correspond to conjunctions of properties specifying the restrictions on a pattern (e.g. a rule like ‘add -ed to the end of a word’ would correspond to a conjunction [-ed and position: word-final]). The goal of the model is to find the best set of rules that explains the data (call it a ‘grammar’) given the set of all potential rules. The choice of the ‘best’ grammar is determined by Bayesian inference which involves calculating probabilities of all rival grammars7 and choosing a grammar that maximizes the likelihood of the data according to Bayes’ rule. The Bayes rule applied to grammars (H) and data (D) is given after this paragraph. It states that the probability of a grammar for some set of data is equal to the probability of the data being generated by this grammar times the (prior) probability of the grammar divided by the cumulative probability of the data given any possible grammar. P(H|D) =

P(D|H) ∗ P(H) P(D)

Frank et al.’s (2008) model implements a simplicity bias by preferring fewer rules (this preference is regulated by a parameter in the model that can be adjusted to make the preference weaker or stronger). An additional bias that this model assumes is a preference for more specific rules which give a tighter fit to the data. (The authors refer to this bias as a principle of ‘minimum generalization’.) In other words, this model implements a typical tension between simplicity and goodness of fit. The authors acknowledge that their model has a serious limitation due to their choice of whole segments (phonemes) as features. If they were to adopt phonological features instead, their model would also be able to capture rules that refer to natural classes of segments (e.g. voiceless obstruent). Nevertheless, this model was able to predict the past tenses of 88.5 per cent of word forms in the testing phase.8 7 In practice, because the hypothesis space is so huge, approximation methods have to be used for these calculations. 8 The predictions for novel items presented during testing were generated by taking the most likely a posteriori inflected form given the most likely hypothesis chosen by the model.

314

katya pertsova

Why was this model able to do so well even without using a realistic set of features? Because many verbs that fall into the same irregular class in English are phonemically similar. So, when the model is tested on a verb it has not seen before, chances are this verb is similar enough to an already discovered category of verbs and so, fits a particular learned rule. If the model did not prefer hypotheses with smaller number of rules, a word-specific rule for each word would be adopted as the best rule. In other words, there would be no generalization. Thus, we see how a simplicity bias (preferring less rules or shorter grammars) drives generalization by taking advantage of structure in the data (phonological similarities among verbs), even if this structure is best captured by statistical tendencies rather than absolute constraints. In contrast to the above model, several non-statistical rule-based models achieve similar results without striving for representational economy (preference for shorter grammars). For instance, the morphological learners described in Albright and Hayes (2002) and in Molnar (2001) both define a generalization procedure that works by iteratively merging properties shared by words undergoing the same structural change. This type of learner first stores word-specific rules (e.g. ‘to get a past tense of poke add -t to the end of the stem’), and then gradually collapses contexts of similar rules. For instance, after a learner has seen the pairs poke—poked and smoke—smoked and formulated word-specific rules for these pairs, it notices the fact that the contexts of the specific rules share a common substring and proposes a more general rule, such as ‘to get a past tense of a stem that ends in [ok] add -t to the end of the stem’. This method leads to the discovery of rules, but it does not entail representational economy. For instance, Albright and Hayes assume that the learner stores all the rules discovered in the learning process, including word-specific rules. This is clearly not the most economical option, but their model ends up closely matching human performance on a past tense wug test. Given that some learners can generalize successfully without minimizing storage space, one might wonder to what extent the storage economy bias is active in human learning, and in the learning of inflection in particular. Perhaps it is not representation size that humans try to minimize, or not only representation size. It is well-known that what may be simple or economical for one component of language may turn out to be complex for some other component. Homonymy presents a simple example: on the one hand, homonymy makes processing less economical since it requires extra time and resources to resolve the ambiguity in the form–meaning mapping (e.g. homonymy presents a disadvantage in lexical categorization tasks); on the other hand, homonymy is economical since it minimizes the number of distinct phonological forms in the lexicon. Thus, the question is: what type of simplicity or economy should the learning procedure try to maximize? Some psycholinguists emphasize the trade-offs between minimizing storage vs. minimizing processing economy, arguing that the best model of human behaviour is the one that achieves a balance between these two competing aims (this view is advocated in the dual-route processing models of Baayen and Schreuder 1968; Caramazza et al. 1988). The conflicting pressures arising from the effects of economy at different levels

machine learning of inflection

315

of grammar may explain why certain types of ambiguity exist in the first place: ambiguity may arise because it is beneficial for some component of language, but at the same time it will be curtailed because it creates a burden for a different component. This kind of economy tug-of-war explanation of statistical tendencies in languages has been proposed for several phenomena ranging from Zipfian distributions of word frequencies to patterns of language change (Zipf 1949; Kirby 2001; Haspelmath 2008). It could be that statistical regularities in morphology are particularly likely to be a product of conflicting economy biases given that morphology serves as an interface connecting several components of grammar, and given that it involves lexical storage introducing additional considerations of lexical access. Yang (2005) is an example of a proposal for how to implement a trade-off between processing and representational economy. Yang presupposes a rule-based morphological system and a serial search processing model in which words and rules are accessed in the order of their frequency. He then proposes a particular metric for the cost of exceptions to rules based on minimizing processing time. The basic idea is that a rule that has exceptions is costly because an application of such a rule requires first scanning all of its exceptions to make sure they do not apply in a given case. Yang suggests that this processing-based cost metric can be used to determine whether a partially regular pattern is most efficiently stored as a rule with a number of exceptions (which themselves could correspond to more narrow rules) or simply as a list of words. Notice that there is a trade-off between the number and frequency of exceptions of a rule and the number of items it explains. A pattern that covers a lot of data with few exceptions will be stored as a rule leading to compression of representation size. On the other hand, an irregular pattern with many exceptions may be better stored as a list of words, especially if these words are frequent and hence easily accessed, leading to savings in processing time. Similar ideas related to defining processing costs can be found in earlier proposals such as Jackendoff (1975). In short, the notion of simplicity is very prominent in computational models of inflection and in learning models in general. Interestingly, the fact that simplicity considerations are active in both production/generation and perception/processing can potentially explain why patterns of ambiguity and irregularity arise in languages and also why they remain restricted. However, it is far from settled what the most helpful measure of simplicity is, and what is the most appropriate way to specify tradeoffs between different types of simplicity. Learning models provide a rigorous way of exploring and testing these issues.

13.5.2 Locality Another prominent bias in morphological and morphophonological learning models regards the notion of locality, the idea that dependencies between morphemes tend to be in some sense adjacent.

316

katya pertsova

The locality bias is well established in phonology as an empirical constraint on phonotactics and phonological alternations (Jensen 1974; Gafos 1996). Physiological and articulatory factors can explain many instances of locality patterns in phonology (i.e. those patterns that are due to co-articulation of adjacent segments or segments of the same kind). So we may wonder whether in other domains, in which such factors are irrelevant, locality effects are less prominent. This does not seem to be the case in morphology. For instance, artificial grammar learning experiments in which subjects learn morphotactic dependencies between nonsense syllables or words show a bias for adjacent dependencies over non-adjacent ones (Gómez 2002; Onnis et al. 2003; Newport and Aslin 2004). Non-adjacent dependencies in such experiments have been shown to be learnable only under certain conditions, either when the amount of training is significantly increased (Vuong et al. 2011), or when the variation in the material intervening between the two dependents is significantly increased (Gómez 2002). These findings raise the question of whether natural language patterns (e.g. nonphonological, contextual allomorphy, and morphotactics) exhibit the same locality bias. While a large scale empirical investigation into this topic is yet to be undertaken, several theoretical proposals (based on small samples of languages) hypothesize specific locality restrictions on contextual allomorphy (Siegel 1978; Adger et al. 2003; CarstairsMcCarthy 2001; Embick 2010). In these proposals conditions on the distance between the dependent morphemes are typically stated in terms of the hierarchical structure of words. For instance, Adger et al. (2003) propose that only sister nodes can be in a conditioning relation. Embick (2010) working from within the framework of Distributed Morphology puts forward a ‘hybrid’ theory of restrictions on allomorphy that involves both the linear notion of locality and the hierarchical cyclic locality (a Minimalism notion). Many learning models of inflection implement an all-or-nothing bias for some type of local dependencies. One interesting case in point is Albright and Hayes’ ‘Minimal Generalization Learner’ for learning morphophonological alternations. Their original learner is biased towards local alternations insofar as its generalization procedure looks for the largest shared material immediately adjacent to the locus of the morphophonological change (Albright and Hayes 2002). In their subsequent work the authors extend this learning procedure to handle non-local dependencies. Interestingly, in doing so they run into a problem. The extended learner applied to Navajo data was able to learn the non-local sibilant harmony generalization (showing up as allomorphy in the perfective prefix), but it also learned eighty-nine other complex generalizations none of which are productive in Navajo, and in fact violate the data not present in the training set (see Albright and Hayes 2006 for discussion and a proposed solution for this problem). The simplest statistical models that incorporate a strong locality bias are the n-gram models. These models rely on frequencies of sequences of n adjacent elements to predict the frequencies of larger strings. Bigram models, a special case of n-gram models

machine learning of inflection

317

where n = 2, are frequently applied to the task of learning sequencing patterns. A proposal by Ryan (2010) incorporates bigram constraints (sequences of two immediately adjacent morphs) within a statistical Maximum Entropy framework (Hayes and Wilson 2008) to learn morphotactics. The goal of Ryan’s learner is to model patterns of free variation in morpheme order which are not due to phonological or semantic factors (the proposal is illustrated using data from Tagalog). This proposal is interesting in that the restrictiveness that comes from using bigram constraints (that is, the assumption that the learner only pays attention to sequencing of immediately adjacent morphemes) proves to derive exactly the typologically attested patterns of free variation such as local context sensitive variation (XYA [*XYB] but YXB [*YXA]), non-transitive patterns (XY, YW, WX), and fails to generate the non-attested ones such as non-local context sensitivity. Moreover, Ryan’s proposal provides an explanation for why free variation might arise in the first place: he shows that when his learner is fed categorical data it can sometimes produce a grammar that generates free variation of the kind that occurs in natural languages. This is another example in which a learning bias explains how certain restricted patterns of ambiguity (here a mismatch between the position within a word and morphosyntactic features typically realized in this position) may arise and be tolerated in languages as a result of learning biases. A locality bias may also be implemented in a gradient rather than categorical fashion. One possibility would be to adopt the hypothesis of Gómez (2002) that learners by default pay attention to adjacent dependencies, but can shift attention to non-adjacent dependencies when adjacent dependencies fail to be sufficiently predictive. This proposal can be implemented in a statistical learning model like the one discussed in the previous section. That is, we can design a model with a trade-off between the goodness of fit and the distance between dependent elements within a rule, so that if the local rules fit the data poorly, competing non-local rules become more probable. The successes of learning models that incorporate a locality bias (and failures of those that do not) support the presence of this bias in learning morphological patterns. However, it is possible that this bias may be ultimately reduced to a preference for patterns that are easier to process. The correlation between processing difficulties (in particular, memory-related difficulties) and long-distance dependencies has been well documented in the literature on syntactic processing (e.g. processing of wh-questions, Phillips et al. 2005; Fiebach et al. 2002). Thus, it is possible that a locality bias could be a particular instance of a simplicity bias discussed earlier.

13.5.3 Conjunctive Bias We have seen that learning allomorphy can be modelled as grouping lexical items into categories (one for each distinct set of allomorphs) based on the properties of these lexemes. In fact, many learning problems can be stated as categorization problems,

318

katya pertsova

in which the task of the learner is to group the stimuli into distinct classes based on their behaviour. Categorization problems have been studied extensively in machine learning and in psychology. In the 1950s and 1960s, psychologists have identified a number of interesting biases when testing how people learn artificially created categories defined by discrete features such as shape, colour, size, and so on. One of the most robust biases they have discovered is a conjunctive bias (Bruner et al. 1956; Haygood and Bourne 1965), or a bias for categories defined in terms of the joint presence of several feature values. This bias is also apparent in grammatical categories of natural languages. For instance, an allomorph class or an inflectional class is typically described as a conjunction rather than a disjunction of properties that condition the allomorphy (e.g. ‘masculine nouns that end in a consonant’ vs. ‘masculine nouns or nouns that end in a consonant’). Similarly, learning of inflectional paradigms can be viewed as learning a number of mostly conjunctive categories. Namely, each affix is a pairing of a phonological string and a featural description of the environments in which this affix is inserted. The featural description most commonly takes the form of a conjunction of features (e.g. [2pl dual masculine]) and much less frequently as disjunction (e.g. [2pl or dual or masculine], equivalent to three homophonous affixes). According to the typological data by Baerman and Brown (2005) from the World Atlas of Language Structures, from a sample of 140 languages that mark subject and person agreement on the verbs, eighty (or 57 per cent) have no syncretism or homonymy in any of the paradigms, which guarantees that every phonologically distinct form can be mapped to a single conjunction of features values. Other typological work on person number marking found that in paradigms with syncretism the most common type of syncretism is the one that can be described by underspecification (Cysouw 2003; Pertsova 2011), which is also a conjunctive type of syncretism. Interestingly, similar results are found in phonology. In fact, for a long time phonologists took for granted that phonological processes affect ‘natural classes’ of segments, that is segments that can be described by a single conjunction of phonological features (e.g. [+stop, +labial]). Mielke’s (2004) dissertation, based on a survey of 561 languages, showed that this was not always the case. Yet, he still found that conjunctive phonological patterns are typologically more frequent, accounting for about 76 per cent of all patterns (classes of segments participating in a phonological rule). The preference for conjunctive categories is not obviously reducible to some notion of simplicity. For instance, simplicity metrics that count the number of symbols in feature representations are unable to distinguish between conjunction and disjunction. From a logical point of view, both operations are also similar: their truth tables contain three instances of one Boolean value, and one instance of the other. One difference between the two types of operations is that conjunction has fewer positive instances in its truth-table compared to disjunction (Table 13.2). One might think that perhaps the fewer the different positive instances of the category, the easier it is to learn. However, this hypothesis is not consistent with a robust finding that inclusive disjunction (three positive instances) is easier to learn than exclusive

machine learning of inflection

319

Table 13.2 Conjunction, inclusive and exclusive disjunction Conjunction

Disjunction

Excl. disjunction (XOR)

F1

F2

F1 ∧ F2

F1

F2

F1 ∨ F2

F1

F2

F1 ⊕ F2

T F T F

T T F F

T F F F

T F T F

T T F F

T T T F

T F T F

T T F F

F T T F

disjunction (two positive instances) (Haygood and Bourne 1965; Gottwald 1971). The truth tables for these categories, in order of their subjective difficulty, appear in Table 13.2. The formal learning theory also does not tell us why simple disjunction should be more difficult than conjunction, although it suggests that conjunctive categories should be easier than categories that can only be represented with complex formulas involving both conjunctions and disjunctions (e.g. disjunctions of conjunctions, or DNF formulas). Specifically, conjunctions are efficiently PAC learnable by a simple algorithm (Valiant 1984), while arbitrary DNF formulas are not (Pitt and Valiant 1988). Similarly, many psychological models of categorization are unable to predict the difference between simple conjunction and disjunction because these models only look for a boundary between examples that belong to a category and those that do not. This strategy cannot in general distinguish between the two types of concepts since disjunctive concepts can be derived from conjunctive concepts by simply switching positive examples to negative and vice versa (assuming all features are binary). An explanation for the conjunctive bias may turn out to lie in the particulars of the generalization strategy used by human learners. Certain linguistic models of morphology are actually consistent with the conjunctive bias because these models assume that the child learns from positive examples alone. There is a very simple learning strategy that can learn conjunctions from positive examples alone, while disjunction is best learned from negative examples. This strategy involves looking for features shared among all the instances belonging to the same category (Siskind 1996; Molnar 2001; Pertsova 2011). Such a procedure automatically favours conjunctive categories because every positive instance of a conjunctive category includes all the properties that are shared by the members of the category at large (possibly together with some irrelevant properties). For instance, if we are learning the category masculine and singular, then all examples of this category will have the property masculine and the property singular, so keeping track of the properties that never change from one example to the next is sufficient to identify all the conjuncts and to rule out the irrelevant properties in the process. On the other hand, when learning a category like masculine or singular, each

320

katya pertsova

example may only have one of the two properties along with many other irrelevant properties. In this latter case, the identification of relevant features becomes much more difficult. Thus, a simple bottom-up memoryless learner whose initial hypothesis is an empty conjunction, and whose subsequent hypothesis upon receiving new data x is derived by intersecting the previous hypothesis with the properties of x, would learn conjunctive categories, but not disjunctive categories. To learn disjunctive categories, the learner has to do some extra work, predicting that such categories should be harder to learn. The learner described in Pertsova (2011) introduces another bias on top of the conjunctive bias, namely a bias towards concepts that can be succinctly described as complements of conjunctive concepts (making a prediction that ‘elsewhere’ type patterns should be the second easiest to learn after the conjunctive patterns). The literature on concept learning includes other kinds of biases related to the category structure that will not be discussed here. The relevance of such biases to the learning of inflection and of other linguistic categories is yet to be systematically investigated (for some steps in this direction see Moreton and Pertsova (2012)). Such an investigation would allow us to systematically compare linguistic and non-linguistic categorization, which could then be used to get more traction on the question of language specialization.

13.6 Concluding Remarks

.............................................................................................................................................................................

The central question of any learning theory is generalization—how can any intelligent system go beyond the available data to make correct predictions about the future? In this chapter, I have attempted to show how machine learning provides insights into this question, and how it touches on several issues of inherent interest to linguists. Perhaps the most central issue is the nature of constraints on human language and the human mind. Machine learning provides a particular angle on this question by allowing us to rigorously explore what constraints on the learner can produce the patterns we find in languages. In terms of acquiring inflection, learning biases help us account for patterns of semi-regularity and limited ambiguity (e.g. allomorphy, syncretism, free variation in affix order). Such learning proposals are particularly valuable in the finite domain of inflection because they make empirical predictions which go beyond the predictions of purely generative or descriptive theories. Generative theories of inflection account for what patterns are acceptable or possible, and typically are hard to distinguish from each other on those grounds alone. On the other hand, learning proposals make predictions for a wider range of phenomena (language acquisition, typological preferences, trends in language change) and hence there are more facts that can be used to choose among these proposals. Because not all representations can be learned in the same way, these facts may also ultimately be used to rule out certain models of grammar.

machine learning of inflection

321

Overall, the main problems that arise for learning inflection, irregularity and mismatches, are in principle solvable (for instance, by the memorizing learner), but linguistically interesting solutions are those that capture the linguistic regularities (categorical or statistical universals) and that predict the trajectory of language acquisition and language change. Such solutions will provide a deeper explanation for the inflectional component of language.

chapter 14 ........................................................................................................

MACHINE TRANSLATION ........................................................................................................

ondrej ˇ bojar

14.1 Introduction

.............................................................................................................................................................................

The goal of machine translation (MT) is easy to define but hard or impossible to achieve in its entirety: to implement a computer program that takes some text in one natural language (the source language) and produces the equivalent in another natural language (the target language). Although MT is a quintessentially linguistic problem, most current MT systems are fuelled by statistics rather than linguistic rules. Since the statistics are based on what has been observed in a corpus, morphological productivity leads to a problem of data sparsity in that many of the word forms will be unobserved or insufficiently observed even in a very large corpus. The problem is compounded by divergences between source and target languages. Divergences arise when one language has morphological exponence of concepts that are not expressed explicitly in the other language. Number, for example, is difficult to translate from Chinese to English because it is not explicitly marked in Chinese. Another type of divergence arises when meanings are expressed by bound morphemes in one language and by free words in another. Hopes for an automatic translation system have been around since the era of John von Neumann and Alan Turing, see Hutchins (2005) or the very optimistic IBM press release in 1954.1 A notable drop in MT research activity followed the ALPAC report (ALPAC 1966; Hutchins 2003), which raised scepticism about the possibility of automatic translation. Since the ALPAC report, the field has recovered its reputation mainly by addressing two shortcomings. First, it was necessary to establish reasonable expectations about MT output and its uses. Fully automatic translation might still be of low quality, but it might be useful for gisting large amounts of material. Higher quality might be required 1

.

324

ondrej ˇ bojar

for the dissemination of technical information such as equipment repair manuals, but this could be achieved if the semantic domain was limited (Reuther 2003; Muegge 2007) or if human post-editors were involved in the process (Koehn and Haddow 2009; Plitt and Masselot 2010; Federico et al. 2012). Secondly, if MT were to be commercially viable and fundable as a research area, it would be necessary to measure progress. Having humans judge MT quality proved to be too slow to support rapid research cycles of evaluation and improvement of MT systems. A giant leap forward was facilitated by the introduction and wide acceptance of automatic metrics of MT quality (Papineni et al. 2002). Despite all the shortcomings of such evaluation methods, as we will see in Section 14.4.5, including a bias towards linguistically superficial statistical methods, the automatic evaluation methods have facilitated large-scale research in MT. The Holy Grail of MT is a system that is fully automatic (without human intervention), domain independent (able to translate a variety of text and speech genres), and high quality. This goal remains elusive. However, many research projects nowadays focus a range of applications that already have satisfied users: •

•

•

free MT services on the Internet: although the output is usually not perfect, it may be useful for understanding or conveying the gist of a text; integration of MT into platforms for computer-assisted translation (CAT), which may include confidence estimation to automatically identify near-perfect segments or mark problematic parts of the output; integration of MT into applications such as cross-lingual information retrieval.

14.2 Machine Translation and Rich Morphology ............................................................................................................................................................................. 14.2.1 History of MT with Rich Morphology Lopez (2008) mentions that a large proportion of MT research deals with translation into English, with the side effect of obscuring deficiencies of the current approaches with respect to morphological complexity. To verify this claim, we analysed papers available in the Association for Computational Linguistics (ACL) Anthology.2 Subject to optical character recognition (OCR) and other conversion errors, our collection as of October 2013 contains 27,015 documents. We selected only those published in 1985 or later that have between 2,000 and 9,000 words to avoid short abstracts or whole workshop proceedings, leaving us with 21,291 papers. About one-quarter of all of these 2

.

machine translation

325

papers mention MT once or more and about 3,500 papers mention MT twice or more with a clear increase in counts since 2007. When searching for language names in the 3,500 papers presumably on MT, we see that 85 per cent of them mention English and about one-quarter of them contains the phrase (in)to English. The second most frequent language is morphologically poor Chinese (mentioned in about a third of the papers). The remaining most studied languages, each appearing in a few hundred MT papers, have differing degrees of morphological complexity (French, German, Czech, Japanese, Spanish, Arabic, Korean, Italian, Dutch, Hindi, and Russian), representing six genetic or typological groups, four of which are Indo-European (Romance, Germanic, Slavic, Indo-Iranian, Japanese/Korean, and Semitic). Many other language families with richer morphology have little coverage, such as Turkic or Uralic (e.g. Turkish is mentioned in about 100 papers, Hungarian in 70 papers and Finnish in 60 papers), or are almost totally absent such as Bantu, Austronesian, Dravidian, and Eskimo-Aleut (used in two parliaments: Canada and Greenland).

14.2.2 Morphological Richness In the field of MT, morphological richness can be measured by examining the number of different tokens in a corpus. We will use the term token to denote an occurrence of a word or punctuation mark in a text. A type of program called tokenizer separates a text into tokens based on various criteria (depending on whether the language uses spaces between words). Punctuation marks also get separated from adjacent words or other punctuation marks, forming tokens on their own. The term vocabulary size refers to the number of distinct token types (referred to as word forms) seen in a given text.3 The word forms cat and cats are two different vocabulary items for us. For our purposes, we define morphological richness only in contrast to another language as the relative vocabulary size compared to the same text in the other language. We will also use the broad term morphologically rich language (MRL) to denote a language richer than English. Consider the book 1984 by George Orwell in the original English and translated to ten other languages. Table 14.1 provides statistics of this book as morphologically annotated in the project MULTEXT-East (Erjavec 2010). Obviously, all the variants of the book tell the same story and the almost identical number of paragraphs and sentences validate the statistics. In contrast, the vocabulary sizes greatly vary across languages, English being one of the least morphologically rich languages and all other ones having 1.82±0.28 times bigger vocabulary on average. The differences in vocabulary size become much less pronounced if we move from distinct word forms to lemmas 3 Perhaps somewhat unintuitively, the term ‘word form’ usually does include punctuation marks and punctuation marks are thus regular entries of the ‘vocabulary’.

326

ondrej ˇ bojar

Table 14.1 MULTEXT-East statistics of the book 1984 by decreasing ‘Lemmas’ count

Slovak Hungarian Polish Czech Estonian Bulgarian Serbian Slovenian Romanian English Persian

Tokens

Vocabulary size in terms of

excl. punct

Word forms

Lemmas

Paragraphs

Sents

84062 80708 79772 79870 75431 86020 89829 90792 101772 104286 95812

20240 20311 21051 19107 17836 16343 18082 17861 15195 9762 11308

10065 10050 9451 9114 8724 8516 8392 8303 7249 7069 6597

1359 1303 1401 1298 1266 1322 1293 1288 1343 1287 1266

6354 6768 6666 6752 6478 6682 6677 6689 6520 6737 6604

(‘base’ or citation forms of words): other languages have on average 1.22±0.15 times more distinct lemmas than English. Morphological richness may be caused by inflection or derivation. However, in MT, and especially in contemporary statistical MT, we are not aware of any issues specific to languages rich by inflection as opposed to languages rich by derivation. What is relevant in MT though, are the differences in morphological richness between the two languages in question. We see four different cases: • • •

•

richer language on the source side (such as Czech→English); richer language on the target side (such as English→Czech); both sides comparably rich and similar in the system of morphological properties as well as syntactic patterns (such as Czech-Slovak), which allows translation by token; both sides comparably rich but more distant in the set of explicitly represented morphological features, word structure, or sentence structure (such as CzechTurkish or German-Finnish), so that translation token by token is no longer linguistically adequate.

Birch et al. (2008) and later Koehn et al. (2009) find, on a large set of European languages, that the target-side vocabulary size is one of the most important reasons for the failure of MT systems of the phrase-based variety (see Section 14.3.3). Finnish was an extreme representative in the experiment, with vocabulary size (in terms of distinct word forms) more than five times larger than English. On the other hand, the source-side richness was not identified as critical. There are two possible explanations. First, the observation may be a side effect of automatic evaluation methods

machine translation

327

(Section 14.4.5), which often require an exact match between the output tokens of the MT system and a human reference translation. With a larger target-side vocabulary, the match is less likely regardless of translation quality. Secondly, MT is actually harder into languages with a larger vocabulary size, because there are more possible outputs to choose from. This choice is further complicated by the fact that less frequent word forms are not adequately observed in a corpus, which negatively affects both the bilingual ‘translation dictionaries’ (Section 14.4.2) and the target-side language models (Section 14.4.4). If both languages are comparably rich and match in the set of morphological properties overt on the surface (of matching words), the system can very successfully decompose translation into independent streams of information, the lexical values and the morphological properties, see also Section 14.5.1. We are not aware of much MT research devoted to rich and morphologically nonmatching pairs of languages aside from a few preliminary attempts for Arabic and French (Hasan and Ney 2008) or Italian (Cettolo et al. 2011) and the broad experiment of Koehn et al. (2009) where, for example, Hungarian–Finnish had very poor results in both directions. Certainly, language pairs excluding English are going to be the next focus of the field and interesting efforts in this respect have already started (Megyesi et al. 2010). On a general note, rich morphology is bound to make the translation harder for simple combinatorics reasons. Having a sentence of n words requires us in principle to consider up to n! word permutations. Modelling morphology explicitly would force us to operate on characters instead of words, considering up to all character permutations of sentences. For m characters in a sentence of n words, m > n and thus m! ≫ n!.

14.3 Anatomy of a Statistical MT System

.............................................................................................................................................................................

Dorr et al. (1998) provide an excellent survey of approaches to machine translation (MT), starting with a list of linguistic problems MT has to handle and including a categorization of rule-based systems depending on the deepest level of linguistic analysis performed by the system (morphology, syntax, or semantics). At that time, there was a division of approaches in MT depending on whether the system consisted mainly of rules written by linguists (rule-based systems) or whether the system consisted mainly of statistics about correspondences between languages learnt from a bilingual corpus (corpus-based systems). Early corpus-based systems that focused on probabilities of translations of individual words did not have the capability to work with linguistic structures such as parse trees and predicate–argument structures. In a more recent survey, Lopez (2008) focuses on approaches to MT that can be called ‘statistical’ (SMT) regardless of the depth of linguistic analysis. The ten years between the surveys saw a great shift towards data-oriented (including statistical) approaches in general. The rivalry between rule-based and corpus-based systems has

328

ondrej ˇ bojar

been softened in that corpus-based systems are now able to work with statistics about structures such as parse trees and developers of rule-based systems are more likely to use a large corpus to discover the necessary rules. The design of any MT system, whether corpus-based or rule-based must designate an approach to the following questions: •

•

•

•

What is the smallest unit of translation (e.g. morpheme, word, string of words, syntactic constituent, or dependency treelet, etc.)? How is an input sentence decomposed into such units (including the selection of some top-scoring subset of all the possible decompositions)? How is each of the source units translated (including the selection of the subset of preferred translations)? How is the output string of words generated from the set of target-side units?

Different choices of the unit of translation and the design decisions for the processes of decomposition, translation of the units, etc. are associated with many different approaches to MT. Translation units usually include morphological information but depending on the system type, the information can be implicit in exact word forms if the system does not analyse them further (e.g. in phrase-based MT, Section 14.3.3), or it can be represented formally using various attributes or node properties and thus detached from the lexical value of the word. Existing systems differ in how far the formalization goes and how easy it is for the system to access the lexical value when handling the morphological attributes or vice versa. Dealing with translation inputs that have not been observed during training is a particular problem in languages with very productive morphology. As most MT systems are currently corpus-based and statistical, we will briefly review the components of a statistical MT system, focusing particularly on the most common type of statistical MT system: a phrase-based system. For other approaches, we just mention some interesting specifics on handling inflection, see Section 14.5.

14.3.1 Learning from Corpora Statistical MT systems are learnt from parallel corpora. One side of the parallel corpus consists of sentences in the source language and the other consists of sentences in the target language. We can assume for now that the source and target language sentences in the corpus line up with each other roughly one-to-one. The parallel corpus is used for learning probabilities, largely by counting things such as how frequently a particular source language word and a particular target language word occur in corresponding sentences; or how frequently a particular target language word follows another particular target language word. During training, probabilities are learnt from the parallel corpus using complex algorithms capable of dealing with large amounts of data, and

machine translation

329

often requiring much computing time. When the system is used for translation, the probabilities learnt in training are applied to a new sentence in order to find its most probable translation. Regardless how large the training corpus is and how complex translation units were chosen (words, parts of words or larger units such as sequences of words, or subtrees in a syntactic representation), we cannot expect to see all possible units in the training corpus. Even if all of the source–target pairs of units are observed, they will not be observed in every possible context. That is, when the system translates a sentence, it is very unlikely that it has seen the precisely same sentence before (unless it is translating a very restricted genre). All practical MT systems therefore include some form of generalization in order to be able to assess the translation equivalence of unseen sentence pairs in the same way as humans do. Breaking up sentences into smaller units is the first vital step to such generalization, but as mentioned, the system will encounter even unseen units. In MRLs, such unseen units are more likely and in order to adequately handle them, the system should generalize further. For statistical systems, the term smoothing is used to refer to this generalization capacity: some function is applied to the observed occurrence counts to ‘smooth out’ the rough edge between a unit never seen and a unit observed once. Unobserved units need some non-zero probability.

14.3.2 MT Pipeline The process of creating an MT system as well as applying it to unseen texts is sometimes referred to as the ‘MT pipeline’. The pipeline originally comes from statistical phrasebased systems (see Section 14.3.3) and it applies well to most data-driven paradigms but at least some of the steps in the pipeline are relevant even for very different MT paradigms such as rule-based MT. The full pipeline (Figure 14.1) includes the following steps: word alignment (i.e. finding corresponding tokens in sentence-parallel training data), extraction of translation units (i.e. automatic construction of a system-specific ‘translation dictionary’), estimation of output fluency (the language model), translation of unseen text, automatic evaluation of MT output, and model optimization (also called tuning; i.e. finding the best internal settings, e.g. weights of individual components of the system). There are four main flavours of relevant data: target-side monolingual texts for language modelling, large parallel texts for extracting translation units and their equivalents, a small held-out parallel text (development set, devset) for model optimization, and finally the input we want to have translated. All the textual data undergo a common preprocessing such as tokenization and tagging. While the preprocessing mostly deals with just technical issues, it easily lends itself to various linguistically motivated tricks, see Section 14.5.2.

330

ondrej ˇ bojar Monolingual

Parallel

Devset

Input

Preprocessing: tokenization, tagging… Word alignment Extraction Language model

Translation model Reordering model

Basic model Parameter/model optimization Optimized model

Translation

fig. 14.1 Typical components of a phrase-based and many other statistical MT systems.

14.3.3 Phrase-based MT (PBMT) Statistical phrase-based MT (PBMT, Koehn et al. (2003)) superseded word-based statistical models (Brown et al. 1988) and has dominated the last decade of MT. One of the reasons is the applicability of the model to any language pair and part of its success also comes from the availability of the complete open-source toolkit Moses (Koehn et al. 2007). The unit of translation in PBMT is a phrase or rather a phrase pair. The term refers simply to a sequence of tokens, completely disregarding any syntactic structure of the sentence. Note that in its basic formulation, PBMT operates with word forms. The phrase pairs translating, for example, the word cat are thus not related to the phrase pairs for cats at all. In Section 14.5.2.1, we describe an attempt to relax this limitation. Without delving into the underlying statistical models, the essence of phrase-based MT can be illustrated by Figure 14.2. The matrix with solid dots illustrates one sentence pair from a large parallel corpus. Each dot corresponds to one automatically estimated point of word alignment. The word alignments serve as the basis for the construction of the critical data resource of PBMT, namely the phrase table. The phrase table serves as the ‘translation dictionary’ for PBMT. It is extracted using just a simple heuristic: all phrases consistent with the word alignment are included. A phrase (pair) is consistent with the alignment if no word is linked to a word outside of the phrase. When a new sentence is to be translated, all decompositions of the sentence into non-overlapping phrases are considered in the search for the best output. Each source phrase is equipped with all its translations as available in the phrase table. The output is gradually constructed from left to right by picking phrase translations for input words that have not yet been covered. Many such partial outputs are constructed and considered in order to select the best output according to a weighted score including the translation probabilities of individual phrases, the ‘language model’ (a model

machine translation . faster even moving ’re they , around time This

331

This time around = Nyní they’re moving = zareagovaly even = dokonce ještě ... = ... This time around, they’re moving = Nyní zareagovaly even faster = dokonce ještě rychleji

zar

N eag yní ova do ly ko nce je ryc ště hle ji .

... = ...

fig. 14.2 Sample word alignment (solid dots) and some of the phrase pairs that are consistent (solid black rectangles) or inconsistent (the grey dashed rectangle) with the English sentence This time around, they’re moving even faster and its Czech counterpart Tentokrát zareagovaly ještˇe rychleji

approximating output fluency, see Section 14.4.4), the number of phrases used, the number of words produced and several other criteria. The space of possible translations is extremely large, so less promising partial outputs are pruned (discarded) early. Pruning can lead to so-called ‘search errors’, that is, situations where a good translation becomes inaccessible because its beginning scored too low and was pruned away. Chang and Collins (2011) show that Moses makes search errors quite frequently but they do not seem to be detrimental to the (automatically estimated) translation quality. Across almost all language pairs, PBMT is relatively successful. Humans may find the output unintelligible or full of errors but it is surprisingly difficult to outperform the PBMT baseline. The reason for this success lies in the ability of reproducing quite long sequences verbatim. In repetitive text domains, this greatly reduces the risk of mishandling known sentences.

14.4 Problems in MT Caused by Rich Morphology ............................................................................................................................................................................. Without adhering to any particular framework, we will describe problems of MT caused by rich morphology in terms of the typical ‘MT pipeline’ as introduced in Section 14.3.2.

14.4.1 Word Alignment Given a corpus of parallel sentences, we need to learn the correspondence between smaller units in the source and target languages. The most apparent units to consider

332

ondrej ˇ bojar

are tokens, although more adequate models aligning morphemes are being studied, see the end of this section. A very powerful statistical model for word alignment is IBM Model 1 (Brown et al. 1993). The model assumes that every word form in the source language (e.g. cat in English) has a probability distribution of correct translations in the target language (e.g. koˇcka in Czech), or possibly no counterpart: cat is a likely translation of koˇcka, while dog is not; the English article the frequently has no Czech counterpart. Given a sentence pair and some estimate of the word-to-word translation probabilities, IBM Model 1 finds the alignment of tokens between the source and target that maximizes the overall translation probability. The Expectation-Maximization algorithm (Dempster et al. 1977) is used to iteratively improve the estimated word translation probabilities and alignment probabilities starting from uniform distribution. Because this is learnt from a corpus with no human intervention, the word translation probabilities (sometimes called lexical translation probabilities) as well as word alignments frequently contain errors. Fortunately, in large datasets, true translations of words dominate the distributions. Rich morphology, and inflection in particular, causes a ‘sparse data’ problem for the task. If a word in both source and target languages does not undergo any inflection, all co-occurrences of the word and its translation contribute to the same entry in the word-to-word dictionary, making the estimate more reliable. Otherwise, for example in Czech–English translation, the model has to learn independently that the Czech zelený (masculine singular, ambiguous between nominative, accusative, and vocative), zeleného (masculine singular, ambiguous between genitive and accusative), zelenou (feminine singular, ambiguous between accusative and instrumental), etc., are all valid translations of the English green. The hardest case are language pairs with both languages rich in inflection: the word-to-word translation lexicon is polluted with up to the full Cartesian product (all combinations) of word forms, for example the German grün, grüner, grünem, etc., vs. the forms of the Czech zelený, see Figure 14.3. Note that the respective morphological features need not be relevant for both languages (e.g. there is no strong or weak inflection in Czech) and even if they are, their values can differ across languages. The problem of sparse data simply arises whenever one Czech word form of a given word corresponds to multiple German word forms and vice versa. The main trouble is not caused by the larger storage requirements of such a polluted dictionary, but rather by the lack of enough training data to support all the combinations with observations. Given for example two sentence pairs such as: (1) green colorless ideas sleep = bezbarvé zelené myšlenky spí (2) I like green pears = mám rád zelené hrušky the model learns that zelené co-occurs with green more often than it co-occurs with pears (twice vs. once in this tiny corpus). However, a new form of the word makes no use of this information in a new sentence pair: (3) I sat under a green tree = sedˇel jsem pod zeleným stromem

machine translation German

Czech

grüner grünen grünem grünen ...

zelený zelený zelenému zelenému ...

Number

Gender

Case

German Inflection

sg sg sg sg ...

m m m m ...

nom acc dat dat ...

strong strong strong weak ...

333

fig. 14.3 A snippet of the Cartesian product of German and Czech morphological variants of the word green

Without further data, the model is completely uncertain whether zeleným means green, tree, sat or any other of the words in the English sentence. Word alignments for rare word forms are established thanks to the assumption that all Czech words map to an English word. If all other Czech words in the sentence are aligned with a high probability, zeleným is likely to align with the remaining green. In combinatorics, this is called the ‘pigeon-hole’ principle: if p pigeons enter h holes in a dovecote and p > h, then at least one hole has to be used by at least two pigeons. The quality of word alignments can be greatly increased by using more advanced models (Brown et al. 1993; Vogel et al. 1996; Liu et al. 2010; Setiawan et al. 2010; V. Graça et al. 2010). Some models have freely available implementations such as GIZA++ (Och and Ney 2003) or a parallelized version MGIZA (Gao and Vogel 2008). However, all of these models treat words as indivisible units. One model (Fraser and Marcu 2007) treats head words and function words differently, allowing only head words to link across languages. The function words are used only for modelling within each separate language. This model would give a promising approach to morphology if bound morphemes could be separated from word stems and treated as function words. The sparse data issue can be mitigated by conflating morphological distinctions in one language that are not overt in the other language. A very rough approximation of lemmas can be sufficient (Corston-Oliver and Gamon 2004; Bojar and Prokopová 2006; Hermjakob 2009), such as the reduction of all word forms to just the first four letters. Such a ‘lemmatizer’ correctly equates tokens like sleep and sleeps under the label ‘slee’, leaving irregularities like slept unhandled. False positives like collapsing the English words envious and environment into a single class ‘envi’ do not cause much harm in the alignment task because the distributional properties of such unrelated words are different and only very few random sentences containing both envious and environment could confuse the model. Bodrumlu et al. (2009) propose a promising model capable of aligning sub-word segments. The model performs well on a small English–Turkish corpus, despite the built-in limitation to 1–1 alignments. Naradowsky and Toutanova (2011) improve alignment quality using source-side linguistic information and automatically segmenting target-side words to morphemes that best match the alignments. Sub-word

334

ondrej ˇ bojar

features (the set of all prefixes and suffixes) in the discriminative alignment model by Dyer et al. (2011a) also allow rich morphology to be handled. Recently, Eyigöz et al. (2013) proposed a two-level alignment model where estimated alignments between words and then between morphemes within words mutually inform each other. Rare words are reported to be aligned more reliably than in other models thanks to their frequent morphemes where the correspondence is stronger.

14.4.2 Extraction of Translation Units The critical component of most MT systems is some dictionary of translation units. If the dictionary is extracted from data, inflection causes the sparse data problem again, as illustrated in Figure 14.3 with the comparison of Czech and German. In statistical systems, entries in the translation dictionary have to be equipped with a probability estimated, for example, as the fractional co-occurrence count (maximum likelihood estimate, MLE): (4)

p(koˇcka|cat) =

occ(koˇcka, cat) occ(cat)

If the dataset were large and repetitive enough, the counts would be relatively high numbers and the fractions would be reliable estimates of the probability. In realworld datasets, such as the proceedings of the European Parliament (Koehn 2005), we will be lucky to see cat co-occurring (and correctly aligned) with koˇcka just a few times. The key problem in the extraction of translation units and estimating their probabilities is in interpreting low co-occurrence counts. If we have seen cat just twice and aligned to koˇtátko (kitten) just once, should we trust it? If the denominator is reasonably high, we can assume a low numerator to be caused by random alignment errors. A low denominator means trouble: the probability of observed translation equivalents is often overestimated, for example the cat being translated as kitten in 50 per cent of cases, while there are salient translations that were not observed and thus have the probability of zero. This is another instance of the sparse data issue and the necessity to smooth the observed counts (Foster et al. 2006; Kuhn et al. 2010).

14.4.3 Translation of Unseen Texts Translation of unseen texts is the heart of MT. The need for generalization capacity concerns all text units down to word forms: the systems need to accept unseen forms and generate novel forms of known words as required by context. Dorr et al. (1998) more or less restrict the issues of target-side generation to cases where the information is not overt in the source language. This affects both the lexical

machine translation

335

value (e.g. the translation of the German können as either know or be able to) as well as morphological properties (e.g. the tense in Chinese-to-English translation). To date, it seems that richer morphology poses problems for MT even in cases where the information is available, simply because the extent of possible choices is too complex to be correctly and fully captured by linguistic rules and/or well covered in available training data. Languages with rich inflection suffer a great deal from a lack of generalization at the word level. Specifically, most SMT systems are completely unable to produce novel word forms or handle unseen words aside from copying them verbatim to the output (which is a reasonable default for proper names in languages with the same script). It is very common, even for very large training data, that a particular form of a word is never seen in the training data. If this ‘out-of-vocabulary’ issue happens on the source side, the system can in principle resort to translating, for example, the lemma instead of the form. Risking an error in, for example, morphological number for nouns, the system increases the chance of having seen the translation unit (now the more general lemma) in the training data and thus producing an acceptable translation; see also the study on error perception by Kirchhoff et al. (2012). Naturally, the generalization capacity is still in the system: it has merely been shifted from the MT engine to the pre-processing step of morphological analysis and/or tagging. If the out-of-vocabulary issue happens on the target side (i.e. the lemma of the target word needed is known, but the required word form is not available), typical SMT systems struggle between producing the most likely translation, that is, the word in an inappropriate form (even though the language model will not give it a high score) and choosing a less salient translation that is available in the form needed by the context of closely neighbouring words.

14.4.4 Language Modelling A very influential component of MT systems is the language model (LM) that is used to select fluent sentences among the many possible candidates. Formally, the language model defines the probability of the sentence being treated as a sequence of tokens. The standard method is called an n-gram language model, because it decomposes the sentence into all (overlapping) sequences of n consecutive tokens, always predicting the last token of this n-gram given the previous n − 1 tokens. While many other sentence decompositions are known to work better, for example those that follow the dependency structure of the sentence (see Chelba and Jelinek 1998; Fox 2002; Popel and Mareˇcek 2010 and the cited works), they are harder to integrate into PBMT (Schwartz et al. 2011). MRLs make the estimation of n-gram LM parameters considerably harder: unless a given n-gram of words is seen in each considered inflection, the model cannot confirm its correctness. Stating that an unseen n-gram is impossible would be too harsh, so LMs are smoothed using one of many proposed methods (Chen and Goodman 1996).

336

ondrej ˇ bojar

The standard technique is to consider a shorter n-gram. So if a particular sentence cannot be confirmed using triples of tokens, only pairs of neighbouring tokens or even individual tokens are checked. Factored LMs (Bilmes and Kirchhoff 2003) allow a linguistically more motivated smoothing path to be specified, for example to preserve the length of the n-gram but consider only, for example, the morphological properties of words ignoring their lexical values or vice versa. So if the sequence big black cats was never seen, hopefully the sequence adj adj noun was. For morphologically rich languages, the reduced model could ensure, for example, agreement in case even for n-grams not seen in the training data, provided that the MT system can actually propose them. The improvement based on this linguistically motivated smoothing in PBMT has so far been unfortunately rather small (Yang and Kirchhoff 2006). Recently, success has been reported in a small data setting with advanced smoothing techniques based on stochastic processes (Okita and Way 2010, 2011) and some of these models are even character-based, allowing them to handle morphemes (Mochihashi et al. 2009). While attempts to model the probability of sentences in a clever way help rather marginally, the brute force of more training data and the standard n-gram LMs is successful and hard to surpass (Brants et al. 2007).

14.4.5 MT Evaluation MT evaluation can serve multiple purposes. If MT is used as a component in a larger process, it can be evaluated extrinsically, using a method relevant for that particular application, for example Krings and Koby (2001) and O’Brien (2011) focus on manual post-editing efforts while Parton and McKeown (2010) aim at cross-lingual question answering. Kirchhoff et al. (2012) offer a rather unusual perspective, evaluating the intuitive perception of MT errors. Using Google translate from English to Spanish followed by manual correction of MT errors, Kirchhoff et al. find that errors in morphology are very common but perceived as far less serious than less frequent errors in word order. Bojar (2011), however, observes for Czech, that errors in word form (including the negation prefix, reversing the meaning of the sentence) can be difficult to spot if the user is presented only with the system output. The more common goal in the MT research community is to measure translation quality during system development in order to check progress or even improve the system automatically (Section 14.4.6). The most useful in this respect are automatic MT evaluation methods. Automatic evaluations are obviously just an approximation of human preferences and do not always match them but they are much faster, cheaper, and also reproducible because they do not depend on an annotator’s subjective criteria and capabilities. The prominence of MT evaluation as such is highlighted by the series of WMT workshops, documented in the works of Koehn and Monz (2006) through

machine translation Source Reference System 1 System 2 Another Reference

337

The earnings on its 10-year bonds are 28.45%. Výnos na jejích 10letých dluhopisech je na 28,45%. Pˇríjmy na své desetileté dluhopisy jsou 28,45%. Pˇríjmy na jeho 10-letých poutech jsou 28.45%. Zisk z jejích 10-letých dluhopis˚u je 28,45%.

fig. 14.4 Exact match of tokens is too rigid for MT evaluation. In this English-to-Czech example, both candidates suffer problems (underlined). Pˇríjmy corresponds to a slightly different meaning of earnings than that used in the original sentence. System 1 selected a wrong case for the translation of on its 10-year bonds, resulting in an ungrammatical but understandable sentence. On the other hand, the lexical choice for bonds by System 2 is completely wrong, the word poutech means handcuffs. In BLEU and other simple evaluation methods, System 1 and System 2 score almost equally. System 1 has preserved the meaning of bonds but due to the different case, the word form is not confirmed by the Reference. As the last line with another human translation illustrates, BLEU would equally penalize a different but correct lexical choice (zisk vs. výnos) and a different morphological variant (dluhopisu˚ ), although it is correct and required by the different preposition (z vs. na).

Callison-Burch et al. (2012) and Bojar et al. (2013b), which regularly include a shared task on automatic MT evaluation and have led to notable improvements. Starting with Callison-Burch et al. (2012), one of the tasks is to automatically predict post-editing effort. The automatic evaluation is usually based on the (monolingual) alignment between the evaluated MT output (also known as the hypothesis) and one or more reference translations. If the target language is rich in inflection, it can happen that the hypothesis and the reference share the content words but, due to a different grammatical relation being chosen, they do not match in form. Simple exact token match as implemented in the most widely used metric BLEU (Papineni et al. 2002) is unable to align such forms and penalizes a significant portion of the hypothesis just as much as it would penalize completely garbage words, see Figure 14.4 for an example. Bojar et al. (2010) report that about one-third of output tokens of English-to-Czech MT systems are not scored by BLEU (because they are not confirmed by the reference) and still do not contain any error based on manual flagging of errors. Having more references helps to mitigate the issue, because they are more likely to confirm the particular forms chosen in the hypothesis. Dreyer and Marcu (2012) propose a technique that captures ‘all’ possible reference translations in a compact data structure and observe that naturally occurring English sentences have billions of meaning-equivalent variations. Bojar et al. (2013a) adapt the technique for languages with richer morphology and morphological agreements that would be cumbersome to ensure in the framework by Dreyer and Marcu. The observations are similar: hundreds of thousands of correct Czech translations can be produced for a single input sentence and having access to them significantly improves the correlation of BLEU

338

ondrej ˇ bojar

with human judgements. However, in their initial stages of development, the techniques are extremely expensive, needing about two hours of manual annotation work per sentence. The task of monolingual word alignment between the reference and the hypothesis has not been studied as much as the bilingual alignment described in Section 14.4.1 (see an open-source implementation by Yao et al. (2013) for some references), but many MT metrics strive to overcome the limitations of exact match. For instance, Tantuˇg et al. (2008) propose a variant of BLEU that considers morphemes and also uses Wordnet similarity to validate root words. The most elaborate handling of the monolingual alignment is implemented in the metric called METEOR (Banerjee and Lavie 2005; Denkowski and Lavie 2010), with several stages: if word forms cannot be matched, lemmas or even Wordnet synonyms or automatic phrase paraphrases (Bannard and Callison-Burch 2005) are considered. Kauchak and Barzilay (2006) transform the problem of alignment for MT evaluation into automatic paraphrasing of the whole reference sentence to better match the output of the system, easing the situation for simple MT metrics. However, neither Kauchak and Barzilay nor Madnani and Dorr (2010) in their survey of paraphrasing methods consider morphologically rich languages. For agglutinative and fusional languages, the issue may also to some extent be solved using letter-BLEU (Yang et al. 2008), a variant of BLEU that is applied on sequences of characters instead of words. Words with matching stems but different affixes would still get at least partial credits.

14.4.6 Model Optimization (Tuning) Current MT system engines always consist of a number of independent components, each associated with a weight. The weights are set automatically in a process called model optimization or tuning to achieve best performance on a heldout set of sentences, the ‘devset’. From a machine learning point of view, this is the actual training of the system. The heldout set of sentences is repeatedly translated and the hypotheses compared to the reference translation using an automatic evaluation metric. As above, in languages rich in inflection, a mismatch in the exact word form is more likely, making the automatic evaluation less reliable. Another issue is that the system should find a balance between a focus on errors in lexical choice and errors in form choice. We are not aware of any study of this balance so far. For the widely used BLEU, this means that both types of errors are considered equally severe, which goes against the findings of Kirchhoff et al. (2012). On the other hand, several studies so far indicate that it is surprisingly hard to come up with a metric that would perform significantly better in model optimization than BLEU (CallisonBurch et al. 2011; Cer et al. 2010).

machine translation

339

Madnani (2010) reports improvements in tuning by adding automatic paraphrases to the reference translation, reducing the sparsity issue. Instead of taking an automatic paraphraser, Dyer et al. (2011b) exploit the fact that their devset was available in multiple languages and report an improvement in the tuning of a German–English system by adding a ‘reference’ as produced by a Spanish–English MT system. Tamchyna et al. (2012) achieve an improvement when tuning English-to-Czech translation on multiple manually created reference translations and Koehn and Haddow (2012b) report better results simply by taking a larger tuning set which also has the positive effect of smoothing out unjustified mismatches with the reference.

14.5 Explicit Handling of Morphology in MT ............................................................................................................................................................................. In Section 14.4, we described in detail how each and every processing step of the general MT pipeline (Section 14.3.2) struggles with rich morphology. In this section, we survey old and contemporary MT systems that do not always quite follow the MT pipeline and, more importantly, include specific modules or procedures to handle morphology.

14.5.1 Shallow Approaches for Closely Related Languages Very close languages such as Czech-Slovak or Spanish-Catalan often differ primarily at the morphological level of representation and share syntactic properties. This lends such language pairs to shallow and direct MT approaches. Word order can be almost preserved and sometimes even the morphological systems are very similar: the same morphological features are overt on the surface and their values can be directly mapped to the other language. An MT system can thus be limited to only performing morphological analysis, including tagging, replacing lemmas and morphological tags in the sequence of tokens, and generating the target-language word form. Depending on the vocabularies of the two languages, the translation of the lemma may need some word-sense disambiguation module, but the morphological properties are often mapped to the target language deterministically. Successful examples of such shallow systems include Apertium (Corbí-Bellot et al. 2005)4 originally developed for the Romance languages ˇ of Spain but gradually extended to cover also less related languages, and Cesílko (Hric et al. 2000; Homola et al. 2009). ˇ Figure 14.5 illustrates the shallow translation by Cesílko from Czech to Slovak including two errors in named entities. Given the nearly one-to-one mapping of 4

.

340 Source Gloss

ondrej ˇ bojar Barack Obama dostane Barack Obama will-get

S. Lemmas Barack Ob S. Tags X@— NNIP7 Output Corrected

dostat VB-S-

Barack Ob dostane Barack Obama dostane

jako as

cˇ tvrtý americký prezident Nobelovu the-fourth American president Nobel

cenu prize

míru of-peace

jako-2 Db—

cˇ tvrtý CrMS1

americký prezident Nobelův cena AAMS1 NNMS1 AUFS4 NNFS4

mír NNIS2

ako ako

štvrtý štvrtý

americký prezident Nobelův americký prezident Nobelovu

mieru mieru

cenu cenu

ˇ fig. 14.5 Sample input and output of Cesílko, including intermediate Czech lemmas and tags (simplified). There are two errors in the output (underlined) and both are caused by insufficient coverage of the morphological dictionary. The name Obama is mis-interpreted as the instrumental (case 7) of the river Ob while the possessive form of Nobel, Nobelův is not covered by the target-side dictionary.

morphological properties, inflection does not pose any unexpected challenge except for unknown words and names in particular.

14.5.2 Morphology in Phrase-based MT This section reviews a range of modifications of the phrase-based model (Section 14.3.3) that treat the input and output in a more adequate way than just atomic word forms.

14.5.2.1 Factored phrase-based MT Koehn and Hoang (2007) introduce an extension of the phrase-based model aimed at explicit handling of morphology (or other features of languages) called the factored phrase-based model. A phrase is no longer a sequence of atomic tokens but rather a sequence of vectors, or ‘factored tokens’. One of the factors is usually the surface form of the word while other factors can represent any information the author of the system deems relevant. The configuration specifies the order in which factors are considered and filled, see Figure 14.6 for an example. The whole preparation of the factored output tokens called ‘translation options’ is performed in an initial phase with no access to neighbouring words. The standard PBMT search follows to pick the best combination of these now richer translation options. It is just the language model that helps to select coherent combinations. Improving target-side morphological coherence is rather easy in factored PBMT models. It is sufficient to introduce additional language models over a subset of output factors, for example part-of-speech or morphological tags (Bojar 2007; Koehn and Hoang 2007; Koehn et al. 2010). This computationally inexpensive benefit, however, concerns primarily the selection of word forms as seen in the parallel training data. Constructing unseen word forms is much more difficult, if we want to avoid a

machine translation

?|gering|? malé|malý|A-pl-acc

?|klein|?

domy|dům|N-pl-acc

?|Haus|? ?|Gebäude|?

?|gering|A-pl-acc-str ?|gering|A-pl-acc-wk ?|klein|A-pl-acc-str ?|klein|A-pl-acc-wk ?|Haus|N-pl-acc ?|Haus|N-pl-dat ?|Gebäude|N-pl-acc ?|Gebäude|N-pl-dat

341

geringe|gering|A-pl-acc-str geringen|gering|A-pl-acc-wk kleine|klein|A-pl-acc-str kleinen|klein|A-pl-acc-wk Häuser|Haus|N-pl-acc Häusern|Haus|N-pl-dat Gebäude|Gebäude|N-pl-acc Gebäuden|Gebäude|N-pl-dat

fig. 14.6 Sample sequence of factored translation of the phrase small houses from Czech (malé domy, ambiguous in case but marked as accusative in the example) to German (kleine Häuser and several other variants). The setup is linguistically motivated: first, Czech lemmas are translated to German lemmas, then Czech morphology is translated to German tags and finally, a German word form is constructed from the lemma and the tag. Later, the best combination of the phrases is selected.

combinatorial explosion of all possible forms; see Section 14.5.2.4 for some promising approaches. Depending on the language pair, data, and exact configuration, factored setups may perform better or run into new problems such as the combinatorial explosion of translation options compared to PBMT (Bojar et al. 2012).

14.5.2.2 Reducing rich source side When the morphologically richer language is on the source side, the translation quality can be increased by stripping the unnecessary details or decomposing complex word forms into separate tokens. Goldwater and McClosky (2005) extend some of the ideas by Yaser et al. (1999) and provide a brief survey of such morphology-stripping methods and evaluate a collection of preprocessing techniques of Czech input when translating to English using a (non-factored) PBMT system. Aside from just reducing Czech word forms to lemmas (optionally equipped with features like number), both Yaser et al. and Goldwater and McClosky also introduce pseudo-words, that is, placeholders that will get translated to English auxiliary words. A minor extension of factored PBMT called ‘alternative decoding paths’ or ‘interpolated back-off’ (Birch et al. 2007; Bojar and Kos 2010; Koehn and Haddow 2012a) allows us to consider both the original rich forms as well as the reduced variants of tokens, taking whichever is easier to use in the given context. The reduced variant is then usually used only as a fallback for unknown forms. Similar results can be achieved with custom models, for example Nießen and Ney (2001, 2004), who introduce a hierarchical translation lexicon where the source word is searched for using gradually less and less specific morphological constraints, or MT techniques capable of handling ambiguous input like confusion networks (Dyer 2007) or lattices (Wuebker and Ney 2012), see Figure 14.7. Nakov and Ng (2011) suggest one more technique to relax the match between the ‘translation dictionary’ as extracted from the training corpus and

342

ondrej ˇ bojar tonbandaufnahme

a0

tonband ton

a1

band

a4

aufnahme a2

auf

nahme

b0

malé malý

b1

domy

b2

dům

a3

fig. 14.7 A sample lattice from Dyer (2009) encoding several segmentation possibilities of the German word Tonbandaufnahme (audio recording) and a sample confusion network encoding lemmatization of the Czech words malé domy (small houses). The original input is displayed as thick edges, the alternative paths serve as fall-back options. Confusion networks are a special case of lattices where all paths reconnect after every edge.

the input. Instead of relying on morphological tools for the source language to produce alternative source tokens, they use automatic paraphrases of words or phrases (Bannard and Callison-Burch 2005). The added benefit is that some divergence in lexical choice can also be accommodated. For agglutinative and compounding languages, the reduction in source-side richness is better achieved by segmenting of complex words. A range of works confirm that such decomposition (or ‘decompounding’ as used in the MT community) is useful for German (Koehn and Knight 2003; Alfonseca et al. 2008; Stymne 2008; Dyer 2009; Hardmeier et al. 2010). The work of de Gispert et al. (2009) decomposes Arabic and Finnish input in many ways, including the unsupervised morphology of Morfessor (Creutz and Lagus 2007) and automatically selects the decomposition that works best for a particular sentence. Nguyen et al. (2010) use a rather different underlying model to achieve the same effect for translating from Arabic and Chinese to English. Virpioja et al. (2010) continue the experiments with German and Czech as source languages; the improvements in translation quality are, however, obtained only when the final outputs are constructed by combining hypotheses from multiple segmentation options.

14.5.2.3 Augmenting poor source side Avramidis and Koehn (2008) enrich English tokens with artificial case markers for nouns and accompanying parts of speech (adjectives, articles, and determiners) and artificial person markers for verbs based on the parse tree. The increase in data sparseness of the source side (due to the richness of the morphology) does not cause much harm, because it is in line with the target language properties (Greek and Czech), where such agreement is required. The underlying SMT model is factored phrase-based using alternative decoding paths to prefer the augmented English tokens but resort to plain tokens if necessary. Yeniterzi and Oflazer (2010) use linguistically motivated rules for pre-processing English when translating to Turkish and gain significant improvements with a small training corpus. They join English tokens to mimic the morphological properties of Turkish, for example concatenating the conjunction if with the verb in the clause or appending the preposition to the head noun of a noun phrase. Assuming that the

machine translation

343

underlying parser does its job well, the features can come from quite distant auxiliary words, effectively allowing some ‘gaps’ in phrases and handling, for example ‘in their . . . economic relations’ as one translation unit (see also Section 14.5.3). A side effect of this transformation is that the number of input tokens drops by 30 per cent, making the number of tokens on both sides similar. Interestingly, a different set of rules aimed at constituent reordering (as opposed to just attaching English auxiliary words to the corresponding content words) did not bring any improvement. Again, the underlying model is factored phrase-based. This time, an additional target-side factor with Turkish morphological tags is used to allow for an additional language model. Yeniterzi and Oflazer also use the alternative decoding paths to resort to the original English source token if the augmented one was not seen in the training data. Ramanathan et al. (2009) improve English–Hindi factored translation in a small data setting by reordering English in a pre-processing step to better match the Hindi SOV word order and replacing English word forms with separate streams of lemmas, suffixes, and automatic semantic relations. English lemmas are translated to Hindi lemmas while English suffixes, and semantic relations are mapped to Hindi suffix or case markers. The final generation step combines Hindi lemma and marker streams to the stream of Hindi word forms. No alternative decoding path (e.g. to ignore the English suffix and relation if their combination is not known) is allowed but the model still outperforms the baseline given the very small training data.

14.5.2.4 Improving generative capacity on the target side Some attempts have been made to increase the capacity for generating new forms. These experiments so far focus on language pairs with only the target side morphologically richer. A rule-based generation component is used by de Gispert et al. (2005) to produce unobserved conjugations of Spanish verbs. Oflazer and El-Kahlout (2007) describe a set of preprocessing techniques for English-to-Turkish translation. Word forms in the target side of the training data are split into morphemes, pseudo-words are added for features such as verb tense (e.g. ‘+vvn’ to indicate passive) and novel word forms are constructed using a separate component from the concatenated stems and pseudowords prior to final scoring. Figure 14.8 illustrates the modified English’ and Turkish’. Producing the final output tokens in the agglutinative language would be very difficult for the standard PBMT, while the split Turkish (e.g. ‘kat +hl +ma’ that deterministically maps to ‘katılma’) allows the simple PBMT model to translate ‘accession’ into the stem ‘kat’ and add the necessary morphemes based on the English auxiliary words. For example, the verbal noun indicator ‘+ma’ probably often co-occurs with the English definite article of deverbal nouns and the model thus learns to translate ‘the’ into ‘+ma’. Note that this can already be seen as a simple variant of the so-called ‘analysis–transfer–synthesis’ approach, see Section 14.5.4.

344

ondrej ˇ bojar

Input

the implementation of the accession partnership will be monitored in the framework of the association agreement .

English’

the implementation1 of the accession2 partnership3 will be monitor4 +vvn in the framework5 of the association6 agreement7 . kat2 +hl +ma ortaklık3 +sh +nhn uygula1 +hn +ma +sh , ortaklık6 anla¸sma7 +sh çerçeve5 +sh +nda izle4 +hn +yacak +dhr .

Turkish’ Output Gloss

katılma ortaklıˇgının uygulanması , ortaklık anla¸sması çerçevesinde izlenecektir . accession partnership’s application , partnership agreement in-framework willbe-followed.

fig. 14.8 Pre-processed English and Turkish (Oflazer and El-Kahlout 2007). For the explanation, content words are boldfaced and the English–Turkish counterparts are co-indexed. A standard PBMT model operates on the modified English’ and Turkish’ representations (without the boldfacing and subscripts).

Several works (Toutanova et al. 2008; Fraser 2009; Bojar and Kos 2010; Fraser et al. 2012) use an intermediate ‘language’ with the target-language word order and lexicon but reduced morphological richness. This artificial language (including the necessary training data) is created by reducing target-side morphological features to a bare minimum. The first step of the translation, performed using a standard phrasebased system, is responsible for most of the transfer and requires parallel data to train. The second step handles the necessary inflection and it can be trained on much larger monolingual data. While successful on small datasets, the benefit of the method diminishes with large datasets. Fraser et al. (2012) use separate models for predicting individual morphological features (case, number, gender, and weak or strong inflection) and are the first to show gains even in a large data setting. Related experiments for the hierarchical model are reported by Weller et al. (2013), see Section 14.5.3 below. Clifton and Sarkar (2011) apply the two-step approach for English-to-Finnish, combining it with ‘segmented translation’ as Oflazer and El-Kahlout (2007), that is, operating the PBMT model on artificial tokens that correspond to (unsupervised) morphemes instead of words. Somewhat related to segmented translation are systems that produce unseen compounds in Germanic languages (Stymne and Cancedda 2011; Fraser et al. 2012). The simplest methods that boast ‘more powerful’ target-side generation use just additional target-side only texts. Sometimes dubbed reverse self-training (Bertoldi and Federico 2009; Bojar and Tamchyna 2011; Lambert et al. 2011), the approach uses an auxiliary MT system trained in the reverse direction (from the morphologically richer language) to translate large monolingual data to the source language. This synthetic parallel corpus is used to train an improved MT system. With some back-off in the reverse translation (e.g. if the form is not known, translate using the lemma), this approach learns to generate word forms never seen in the original parallel data.

machine translation

345

14.5.3 Hierarchical and Surface-Syntactic MT Presented as an extension of the phrase-based model, the hierarchical phrase-based model (Chiang 2005, 2007) correctly handles the hierarchical structure of sentences. Phrases (in the non-linguistic sense of token sequences) can now contain gaps where other phrases fit. In the hierarchical model, this composition is not restricted by the type of the phrase in any way; in a (surface) syntactic model, phrases and gaps are labelled with non-terminals that have to match, formally making a synchronous context-free grammar (Chiang and Knight 2006). If the non-terminal labels and the recursive structure come from a treebank, we have a truly syntax-based translation. The motivation for such a model from the linguistic point of view is obvious and includes, for example the chance to capture long distance dependencies between words (e.g. agreement in some morphological feature). The non-terminals in hierarchical model can be used to encode not only syntax but any other latent feature (information not overt on the surface of source or target languages), for example Baker et al. (2012) use this for better handling of modality and negation. The additional constraint requiring non-terminals to match has to be introduced with great caution. An option to resort to non-matching phrase has to be allowed in order to preserve the performance of the plain hierarchical model. Otherwise, the rigid syntactic model effectively reduces the available training data by disabling non-matching phrases (Bojar and Hajiˇc 2008; Chiang 2010). Across language pairs, hierarchical translation has not quite outperformed PBMT. The added complexity and constraints of the structure do not always pay off. Since the search for a robust hierarchical model is still in progress, relatively few people have tried to focus specifically on morphology in this model. Williams and Koehn (2011) are probably the only ones who attempt to formally capture agreement constraints in the hierarchical model using proper unification instead of simple identity of non-terminals. Weller et al. (2013) use the hierarchical model as the basis for their two-step approach (Section 14.5.2.4) and report gains when predicting case for German noun phrases using a range of features up to subcategorization frames.

14.5.4 Systems Following Analysis–Transfer–Synthesis Sequence Approaches to MT that follow the analysis–transfer–synthesis sequence (Vauquois 1975; Vauquois and Boitet 1985) include separate components for handling morphology. The lemma and the various morphological features of words are separated during the analysis phase. The transfer can thus handle morphological and lexical divergence separately. Finally, a fully-fledged morphological synthesis can produce forms never seen in the training data. The exact specification of the morphological component and the set of observed features is system-dependent.

346

ondrej ˇ bojar

Unfortunately, not many such systems have made it up to the ‘production’ state. We are aware of two commercial systems: MT-MSR (Richardson et al. 2001), later discontinued, and Lucy Technologies (Wolf et al. 2010), originally based on the METAL system (Bennett and Slocum 1985), and a few research systems. A moderately-sized MT system for the translation from Norwegian to English was achieved in the LOGON project (Bond et al. 2005; Oepen et al. 2007; Bond et al. 2011) using an interesting combination of Norwegian analysis in Lexical Functional Grammar (LFG, Bresnan 2001), transfer in Minimal Recursion Semantics (Copestake et al. 1995), and English generation using Head-driven Phrase Structure Grammar (HPSG, Pollard and Sag 1994). Meaning-Text Theory (MTT, Mel’ˇcuk 1988; Kahane 2003), is exploited in the MT system ETAP-3 (Apresjan et al. 2003; Boguslavsky et al. 2004).5 The transfer happens at a normalized syntactic representation (NormS), slightly above surface syntax. TectoMT (Žabokrtský et al. 2008; Dušek et al. 2012)6 is based on the tectogrammatical layer (t-layer) of linguistic representation (Sgall et al. 1986) also known from Prague dependency treebanks (Hajiˇc et al. 2006; Hajiˇc et al. 2012) and translates only from English to Czech. Compared to MTT, the t-layer is less semantic and stops at the level of lemmas and morphosyntactic realizations of relations between words (e.g. capturing preposition and case instead of some deep syntactic or semantic role type). Transfer at the t-layer thus has to handle context-dependent choices in case markers and other morphosyntactic markers. TectoMT uses a statistical model for this (Žabokrtský et al. 2010), circumventing the need to manually encode lexical functions or grammatical rules. Bojar and Hajiˇc (2008) and Bojar and Týnovský (2009) document the additional problems that such a complex processing pipeline creates compared to shallow or direct approaches. This explains to some extent why deep approaches have not surpassed the performance of simpler models yet, despite the obvious improvement in linguistic adequacy. TectoMT remains probably the only such deep system that performs reasonably well in the broad domain of news (Callison-Burch et al. 2010).

14.5.5 System Combination and Corrective Approaches So-called ‘system combination’ techniques can be used to benefit from the strengths of diverse types of MT systems such as lexical choice of large-data PBMT and better grammar and unseen word forms produced by more syntactic and/or rule-based systems. The seminal work on MT system combination (Matusov et al. 2008) follows the idea originally introduced for automatic speech recognition (Fiscus 1997). This one and subsequent variations, for example Heafield and Lavie (2010), are generally based 5 6

. .

machine translation

347

on some voting for individual tokens: if many baseline systems produce a particular word form, then it should probably appear in the combined output. None of the system combination techniques addresses inflection explicitly, all rely on the standard n-gram LMs (Section 14.4.4). It is thus again only the size of the dataset that is supposed to bring some guarantee of grammaticality. Rosa et al. (2012) take a different approach and implement Depfix, a rule-based system that fixes (primarily morphological) errors in a baseline MT output given automatic syntactic analyses of both the source and the hypothesis. A complex ensemble of the deep syntactic TectoMT (Section 14.5.4), the phrase-based Moses in a factored setup (Section 14.5.2.1) and Depfix for final correction was the best performing English-to-Czech system in WMT13 shared task (Bojar et al. 2013c). Aside from various changes in verb conjugation, the most reliable automatic correction was the re-introduction of lost negation.

14.6 Summary

.............................................................................................................................................................................

The inflection and morphological richness of source and/or target languages introduce extra complexity to almost every processing step of machine translation systems, starting from the alignment of words in parallel texts up to MT system evaluation. In the history of machine translation, the relatively morphologically poor English has long been the most common target language. Specific handling of inflection and rich morphology has thus received proper attention only rather recently. The currently prevailing data-driven approaches still have a long way to go until they adequately capture and apply the necessary morphological generalizations. The former opponents, rule-based vs. statistical methods, have grown very close to each other and we expect even further convergence on this journey to correct generalizations. It is primarily the underlying material that forms the interest and focus of research. We therefore expect many influential discoveries to arise from the study of machine translation between morphologically rich but divergent languages, a sector where the explorations have barely started. While the quantity of parallel and monolingual texts suitable for the training of MT systems is growing every minute, fine-grained models of inflection (and word formation) remain a needed component of general-purpose MT systems, because new word forms are constantly being created. Sooner or later such models will be designed and become a part of the standard MT pipeline.

part v ........................................................................................................

PSYCHOLINGUISTICS ........................................................................................................

chapter 15 ........................................................................................................

INFLECTIONAL MORPHOLOGY IN LANGUAGE ACQUISITION ........................................................................................................

sabine stoll

15.1 Introduction

.............................................................................................................................................................................

Languages vary to an extreme degree in their structures, and one of the major tasks of language acquisition research is to explain how children can learn any of these structures. Recent research has shown that universals play a much less important role in linguistic theory and empirical research than previously assumed (Bickel 2007; Evans and Levinson 2009). Thus, the question arises what strategies or learning mechanisms do children rely on in learning a language, in our case morphology? The present chapter focuses on two main issues: (i) the interaction between learning and the language-specific, grammatical build-up of languages from a cross-linguistic point of view and (ii) the strategies children choose in learning different morphological systems, such as item-based or rule-based learning. We can expect that different morphological structures pose different challenges to the learner (Slobin 1985b; Peters 1997) and that these structures influence both the speed and strategy of learning. Two types of variation are relevant for research on the acquisition of morphology. First, there is cross-linguistic variation in the expression of morphology. Languages range from having hardly any morphology to exhibiting extreme degrees of polysynthesis, and acquisition theories need to account for the acquisition of this whole range. Second, there is individual variation on several levels, such as the input children receive, their speed of learning and the approaches they take in learning the structures of their native language (Lieven et al. 1992; Bates et al. 1995).

352

sabine stoll

The first major task of the language learning child is to segment words into their morphological structure. This requires a number of perceptional, computational, and memory skills. The features most relevant for the acquisition of morphology are probably the number and kind of morphemes a child has to learn.1 The main characteristics of morphology that are relevant for segmentation and abstraction are: (i) degree of synthesis; (ii) phonological fusion; (iii) presence or absence of inflectional classes; (iv) homonymy of forms; (v) presence or absence of a default and number of exceptions; (vi) semantic fusion; (vii) phonological shape of morphemes and allomorphs. In addition, it is presumably important how morpheme boundaries are realized, that is, whether they are opaque or easily detectable. For this question the kind of allomorphy available is relevant as well, that is, whether allomorphs are predictable or suppletive (Peters 1997). The task of recognizing morpheme boundaries in a language like Mandarin with few but easily detectible grammatical morphemes differs strongly from the task in a language like Navajo (Dené-Yeniseian, North America) or Mohawk (Iroquoian, North America), where morpheme boundaries are difficult to detect. Peters hypothesizes that ‘probably the hardest morphemes of all are those that are unstressed, are word-internal, straddle syllable boundaries, and have opaque functions.’ (Peters 1997: 182). According to this hypothesis, children learning a morphologically complex language will become productive later than children learning a language with little or extremely regular morphology. Thus, we can expect the quantity of morphemes and their morphophonological character to interact in learning morphology. These interactions then are relevant for frequency effects in the input, which again play a role in processing language. Frequency patterns of word forms are very different in analytic and synthetic or even polysynthetic languages and these different frequency patterns are likely to have have an impact on learning (see also Pertsova, this volume on machine translation). Thus, children learning languages with little morphology get more repetitions of the same word form in different contexts than children who learn a language with elaborate morphology. The generalization direction in an analytic language like English thus proceeds from one word form to many contexts, and as a result, the child encounters many repetitions of the same word form. Such generalizations would not work in a synthetic language with elaborate morphology which exhibits separate word forms for different contexts. Thus, a child learning a synthetic language with a lot of inflectional morphology has to remember many forms in specific situations and generalization from one word form to a pattern is a more tedious and longer process than in an analytic language. However, one could also expect that the ubiquitous presence of morphology in a language could have the opposite effect: that children who learn synthetic languages have to focus more on morphology right from the beginning of their linguistic life to be able to understand and express anything. This might ease and even speed up the process of learning morphology (e.g. Dressler 2007). This line of thinking is supported 1

I make no theoretical comments beyond this.

inflectional morphology in language acquisition

353

by experiments which show that brief auditory exposure to any kind of repeated presentation of artificial linguistic material is enough for prelinguistic children to generalize statistical patterns (Saffran et al. 1996; Saffran 2002; Gómez and Gerken 1999). Thus, early generalizations are found in processing, where no meaning is involved. However, it is disputed whether such immediate generalizations also apply to early production, for which the core component actually is the transmission of meaning, or whether children in early acquisition are instead more conservative and rely on itemspecific learning, that is, they first restrict their uses of grammatical forms to specific lexical items and only later on in development generalize to other items (e.g. Tomasello 1992; Lieven et al. 1997). Thus, the main question we focus on in this chapter is which strategies or acquisition mechanisms are at work when children cope with the different task of the diverse morphologies found in the languages of the world. As with other branches of linguistics, research studying the acquisition of morphology has been divided by the controversy of whether universal pathways of learning exist or whether language-specific structures set the pace in acquisition (for a detailed theoretical discussion of these two approaches see Ambridge and Lieven 2011). Universalist approaches to learning usually assume rule-based learning, even though there is no intrinsic relation between innateness and the assumption of rule-based learning. Rule-based approaches—also known as dual-mechanism models—generally assume a clearcut distinction between grammar on the one hand, which they claim is learned by innate rules and the lexicon on the other hand, which is assumed to be rote learned (e.g. Pinker 1984; Pinker and Prince 1990; Clahsen 2006). Thus, in rule-based accounts it is assumed that there is one default, this default accounts for regular forms which are processed by rules, whereas irregular forms are stored in the lexicon and learned by rote. Constructivist approaches, also known as emergentist or usage-based approaches (for an overview see Behrens 2009), by contrast, assume that there is inductive, schema-based learning of language-specific structures over which a language learner then generalizes and thereby builds up a grammar from bottom to top. The learning process is taken to rely on small-scale associations called schemas, which are based on phonological and semantic associations. Rules in this approach are assumed to be secondary, abstract generalizations over these schemas, that is, they are the results of pattern detection, which are then generalized to other instances (e.g. Tomasello 2003). Thus, the basic idea behind schema-based learning is that children learn lexical specific constructions by heart and then substitute parts of them step-by-step. In this way they build up their grammar, which is nothing more than a construction inventory. In contrast to universalist approaches, in schema-based learning linguistic categories and principles are not considered to be domain-specific and innate, but are rather hypothesized to be based on general cognitive abilities. These theories, which usually take a single-mechanism approach to learning, claim that there is no categorical distinction between grammar and the lexicon (e.g. Bybee 1985; Tomasello 2003) and hence there is no difference in the learning mechanisms of the two. Both rule-based and constructivist theories need to account for the acquisition of morphology outside of the usually cited examples of learning the English past tense

354

sabine stoll

or the English plural. The past tense, in a language like English indeed has a clear default with a restricted number of exceptions. Thus, in English, the acquisition of the past tense might indeed be explained by rule-based learning. In other languages, where no clearcut default can be extracted such an account would need to postulate a number of rules and then in addition decision mechanisms to choose the appropriate rule need to be formulated. In languages with large numbers of conjugational classes as, for instance, in Dumi (Sino-Tibetan, Eastern Nepal), which has five intransitive conjugations and eleven transitive conjugations which distinguish the different patterns of stem alternations (Van Driem 1993: 90) or in languages with a large number of idiosyncratic forms such as for instance Mohawk (Iroquoian, USA, Canada), where subject and object agreement is marked mostly by portmanteau morphemes (Mithun 2001), the child will not be very successful in approaching morphology by applying one default rule; rather the distribution of patterns or better schemas across the lexicon need to be determined. Constructivist approaches would assume that low-scale analogy and schematization processes are responsible for learning morphology in such a system without a clearcut default rule. One of the main challenges in comparing the acquisition of morphology is to deal methodologically with the extreme variation in form and function of morphological categories across languages. This challenge will be addressed in the first section of this chapter, focusing on some new methods to measure development and productivity that have been recently introduced in the field. Subsequently, we turn to the role of morphology in the acquisition of the most prominent parts of speech in early acquisition, that is, nouns and verbs, and thereby focus on those inflectional categories that are the most widely studied in acquisition research, that is, case and number for nouns and tense/aspect and voice for verbs.

15.2 Development from a Cross-linguistic Perspective ............................................................................................................................................................................. 15.2.1 How to Determine Productivity? One of the most pressing questions in research on morphological acquisition is how to determine when children become productive with a specific feature. Two types of productivity are usually discussed in the literature but they are not always distinguished properly. First, there is morphological productivity in the sense of the ability to apply a feature independently of a specific context and independent of specific lexical items, usually understood as the application of a morphological rule to an unknown stem. A now widespread method of studying productivity are experimental set-ups that use nonce words and test whether children apply the same rules that an adult native

inflectional morphology in language acquisition

355

speaker would apply and what overgeneralization errors they make. The classic experiment introducing this technique was Berko’s famous wug test, an experiment with nonce words, studying English preschooler’s ability to use the plural and the past tense (Berko Gleason 1958). For measuring productivity in longitudinal data, Brown (1973) used as a criterion for productivity the ‘correct use’ of a variable in 90 per cent of the obligatory contexts in three successive samples. Another widespread criterion for productivity is the use of a specific marker with different items, for example Shirai et al. (1998) require an aspectual marker to occur with at least five different verbs within one month of recording to call the marker productive. Other researchers apply a combination of criteria, such as for instance Ketrez and Aksu-Koç (2009) in a study of the acquisition of noun morphology in Turkish. They use a combination of four criteria to determine productive behaviour: ‘(a) a morpheme should occur with at least 2 nouns; (b) the nouns occur in at least two contrasting forms; (c) grammatical morphemes are used more often correctly than not correctly; (d) the instances occur in 2 successive samples.’ (Ketrez and Aksu-Koç 2009: 21). These criteria, however, strongly interact with the frequency of the feature investigated and individual contexts of recording, and further they are completely arbitrary, non-statistically determined cut-off points. Second, the term productivity is often used to characterize the behaviour of the child as adult-like. To avoid confusion, I distinguish the two types terminologically and use the term productivity for the ability to generalize and the term proficiency for adult-like behaviour (see Stoll et al. 2012). In recent research more quantitative methods have been proposed to characterize the development of inflectional diversity and to determine proficiency, that is, adult-like behaviour. Richards and Malvern (2004) showed that type/token ratios, which have been widely applied in acquisition research, strongly depend on sample size. The bigger the sample the smaller is the type/token ratio, independent of development. As a consequence of this observation, they developed a measurement which includes a modified type/token ratio. A similar measure has been proposed by Xanthos and Gillis (2010) investigating the mean size of paradigm. Both approaches use a bootstrapping method in processing the corpus as several different subcorpora, that is, they extract several subcorpora from the corpus with a number of samples differing in sample size, which they then subsequently analyse. The great advantage of these two methods is that they take sample size into account, which is important in the study of longitudinal, conversational data with a variable number of utterances per recording. Recently a number of further measures have been proposed to compare the behaviour of children to those of their caretakers. The basic idea is to statistically compare the child’s inflectional flexibility to that of their surrounding adults. Aguado-Orea (2004), in a study of productivity of verb inflection in child Spanish, chose matched samples of children and adults (as parts of a longitudinal study) at different points of development. He counted a child’s average number of inflections per verb and compared it to the average of the same verb in these carefully matched samples of the caregivers. This shows whether children use the same amount of inflections as adults

356

sabine stoll

do with carefully matched verbs. Stoll et al. (2012) in a study on the acquisition of Chintang (Sino-Tibetan, Eastern Nepal) nouns and verbs investigated the development of morphology in applying relative information entropy. This measure of relative entropy compares the inflectional proficiency of the target children to that of the surrounding adults. Entropy is used to measure how predictable the occurrence of a given form of a paradigm is. The more predictable the use of specific forms, the less flexible and proficient the speaker is. Thus, we can expect that at first children are restricted in the forms they use and later on they will develop toward adult distributions. In Chintang all stems behave the same and there are no conjugation classes. Consequently, in this study all forms could be subsumed under an abstract stem template and the distribution of forms irrespective of the verb stem in the use of a specific child and her corresponding adult data were compared. Like the measurement of mean size of paradigm by Xanthos and Gillis (2010), entropy controls for the total frequencies used by the speakers. Krajewski et al. (2012) use a similar method, partly based on AguadoOrea (2004), however in this study the goal is to find out how flexible children are in the use of inflections in context, that is, in the lexical items surrounding the inflectional form. The goal is not to find out how the child differs in the amount of inflections used but whether there is a difference in the use of the same inflections. Before counting the entropy of the speakers, they thus first extract the inflectional endings used by the child and then randomly choose the same number of utterances with the same inflectional endings from the adult corpus. Subsequently, they examine the contexts in which these inflectional forms occur and compare the flexibility of use between adults and children. All these methods allow the comparison of the performance of a single child to that of her surrounding speech. It is well-known that children within a language show strong variation in the speed and also in the ways they acquire their native language (Lieven et al. 1992; Bates et al. 1994; Fenson et al. 1994). Although large-scale experiments can control for individual variation, in longitudinal studies, which can usually only include a small number of participants, individual variation is more difficult to control for. Recent research by Dabrowska ˛ (2008) has claimed that individual variation extends also to adult competence in showing that productivity with Polish case inflections is subject to individual variation in adults. This entails that competence with grammatical categories is not a homogeneous feature of adult native speakers. Consequently individual variation needs to be included in models of language acquisition and adult speech. Thus, instead of making gross generalizations over a large number of unknown variables, an incremental approach seems to be the best choice in order to account for individual variation in longitudinal data. This implies that we first analyse each child individually and only later, when we know more about her development, we compare it to the development of the other children. Thus, as a first step one determines the individual development of each child per language and compares his/her performance to that of the surrounding caretakers. To find out whether the resultant pattern of development can be generalized, patterns of various children need

inflectional morphology in language acquisition

357

to be compared. If there is a common development in the children and the behaviour of the adults is similar to each other, one could consider pooling the individual dyads and start cross-linguistic comparisons. This same incremental procedure needs to be applied in the other language/s of the sample and only then the pathways of development can be compared across languages. Cross-linguistic comparisons are probably the biggest challenge because of the heterogeneity of categories in form, function, and occurrence and the variation encountered in individual children.

15.2.2 Methodological Challenges in Cross-linguistic Comparisons Comparative language acquisition research combines the complexities of typological and acquisitional research. It faces both the comparability problem known from typology—that is, it has to consider the extreme variation of grammatical categories on a number of levels, such as morphological complexity, semantic extension, frequency of occurrence, transparency of expression, saliency in position, polyfunctionality, obligatoriness, etc.—and it has to face the individual variation problem of acquisition research in including individual variation in development and adult competence. However, to take all these variables into account a lot of data of different languages are necessary and this is perhaps the biggest challenge to comparative language acquisition research. There is only a severely limited sample of languages for which we have longitudinal data available and this small sample is very much biased towards the Indo-European languages of Western Europe. These data restrictions limit generalizations about acquisition beyond individual languages significantly (Stoll 2009). Thus, a key issue for comparative studies is to include a wide variety of different morphological structures. We know from a number of studies on the acquisition of verb inflection that the type of morphological marking and the degree of morphological complexity is a key component for the speed of acquisition (Aksu-Koç and Slobin 1985; Laaha and Gillis 2007; Stoll et al. 2012). So far there is no commonly accepted method for measuring and comparing complexity across languages. One approach is to rely on the traditional typological classification of ‘weakly inflecting languages’ (German, Dutch, English, and French), ‘strongly inflecting languages’ (Croatian, Russian, Greek) and ‘agglutinating languages’ (Turkish, Finnish, Yukagir) (Laaha and Gillis 2007). This classification, however, has been strongly criticized from a theoretical point of view for being based on confused and conflicting criteria (Bickel and Nichols 2007; Haspelmath 2009). The availability of large and fine-grained databases allows for new bottom-up approaches. Stoll and Bickel (2013) use a dozen typological variables expected to have an impact on acquisition, concerning such issues as the presence and nature of agreement and case marking, degrees of synthesis, polyexponence, and inflectional compactness of categories (e.g. categories expressed cumulatively in one affix or distributed over the word),

358

sabine stoll

syncretism, the existence of inflectional class, etc., all cross-classified by word order because the position of morphology in the clause is known to be relevant for acquisition and processing. Using clustering algorithms, the authors find that languages do not fall into natural clusters of varying complexity. At the same time, however, the algorithms also suggest a useful heuristic for sampling languages that are maximally distinct across the relevant variables.

15.2.3 The Acquisition of Complex Morphologies Languages differ enormously in the complexity of their morphological system(s) and it is one of the biggest puzzles of acquisition how children manage to cope with this diversity. Surprisingly, this field is largely underresearched, and there are no large-scale comparative studies on the influence of different morphological systems on the acquisition process. Traditionally research concentrates on the acquisition of parts of speech with a focus on morphology, such as on the acquisition of verbal or noun paradigms. Results of this research indicate that morphological complexity has an impact on early distributional patterns of early child speech. A number of acquisition studies on typologically unrelated languages such as Mandarin (Sino-Tibetan, China, Tardif 1996), Korean (Choi and Gopnik 1993, 1995), Ngas (Afro-Asiatic, Nigeria, Childers et al. 2007) and Tzeltal (Mayan, Mexico, Brown 1998) suggest that morphological complexity is an important factor for an early noun or verb preference in children. In most languages nouns are morphologically less complex than verbs. A further factor that has been in the focus of research is input frequency. In Mandarin, for instance, verbs are more frequent than nouns, In naturalistic studies (Tardif 1996; Tardif et al. 1997) an early preference of verbs was found in the data of Mandarin children which closely matches the input they receive. In Mandarin, verbs usually occur in salient position, that is, since subjects and objects are frequently dropped, they often occur in utterance final or utterance initial position or even completely on their own. Further, verbs show no morphology thus children do not have to learn several forms depending on, for example, person or number or tense. Chintang, another Sino-Tibetan language, also exhibits strong argument ellipsis, and has verbs in salient last position. However, in contrast to Mandarin, in this language children are confronted with an extremely high degree of verbal complexity. This verbal complexity is hypothesized to play a role why children in early acquisition do not match the adult distributions (Stoll et al. 2012). Adults show a very low noun to verb ratio but children in early development up to age 3 focus on nouns first, which are morphologically much simpler than verbs. Only later on in development when children become more proficient with verb morphology do they also start matching the noun to verb ratio of the adult language. Thus, it seems that a number of language specific structural factors are involved in learning parts of speech and morphological complexity is one of them.

inflectional morphology in language acquisition

359

In a language like Chintang with extremely complex verb morphology children seem to avoid verbs at first. From the literature we know other avoidance strategies of morphology in early acquisition. We have evidence for cross-linguistic variation in whether and how frequently children omit inflections in early acquisition. Children learning English, German, Dutch, Swedish, Danish, French, Russian, and some other languages were found to omit inflections in verbs and instead use only infinitives. Evidence from English is inconclusive whether these forms are roots or infinitives (for a summary see Hoekstra and Hyams 1998). Results from languages such as German (Poeppel and Wexler 1993) and Russian (Brun et al. 1999; Gagarina 2002) has led researchers to argue that these forms are infinitives (for discussions about how these ‘optional infinitives’ are explained by different theories, see Rizzi 1994; Wexler 1994; Aguado-Orea and Pine 2002). Root forms, but not in the form of infinitives, have been found in other languages such as Navajo (Na-Dené, USA) and Quechua (South America) both of which show extensive verb morphology (Courtney and Saville-Troike 2002). In Quechua the stem is in initial position followed by a number of different affixes. In this respect Navajo is the mirror image of Quechua, with a number of prefixes followed by the stem. In both languages Courtney and Saville-Troike found root forms, which children never heard in the input and which do not correspond to the infinitive in these languages. These results suggest that bare roots or infinitives are a language specific feature of some languages. In German, Russian, and Dutch the occurrence of ubiquitous infinitives might be due to the frequency of modal constructions and complex verbs, together with related features of perceptual saliency, that is, they occur in utterance final position (Freudenthal et al. 2010). The reason for the occurrence of root forms might be different for languages with different grammatical features such as Quechua and Navajo. However, the occurrence of roots or infinitives in early child language is not a universal phenomenon; children, for instance learning Inuktitut (Eskimo-Aleut), a polysynthetic language spoken in northeastern Canada (Crago and Allen 2001) or Chintang (Stoll et al. 2012), do not seem to go through a phase using either infinitives or roots without inflections. To resolve under what circumstances children avoid inflections more comparative research is needed and a much larger database of unrelated languages with a variety of morphological structures needs to be included in the sample. Another issue that has not yet been approached in much detail is the question of how children learn complex morphological systems, such as the large paradigms found in polysynthetic languages. It is far from clear which strategies children use to learn, for example, a complex verb paradigm as found in Chintang, where each verb has at least 1,849 forms (Stoll et al. 2012). What we do know, however, is that children take quite some time to gain the same proficiency as adults do. To test this, measures like entropy are of great help, since they allow us to estimate the proficiency of a specific linguistic feature in relation to adult performance. This measure allows an overview of

360

sabine stoll

performance based on large data samples, instead of relying on mere proportions as has been common practice in the field so far. In the following section, the focus will be on what we know about the acquisition of some inflectional categories in a number of different languages for which we have data.

15.2.4 The Acquisition of some Nominal Categories 15.2.4.1 Number In most languages the plural on nouns is marked whereas the singular is the unmarked default (Greenberg 1966). Thus, most acquisition studies have concentrated on plural marking. A large-scale study by Marcus (1995) based on ten longitudinal studies taken from the Child Language Data Exchange System (CHILDES, ) initiated by Snow and MacWhinney, found that English children rarely overextend regular forms to irregular stem forms. Marcus counted the mean of all overgeneralizations (8.5 per cent) and concluded that these overgeneralizations are negligible. For past tense, he found a similar pattern. From that he concluded that inflectional learning is rule-based, and that once the children have extracted the rule, they operate it as a default. However, it is far from obvious how to determine whether a certain number of overgeneralizations is negligible or not. After all, 8.5 per cent is not such a small number, particularly if for all children there are numerous high pockets of errors for specific recording sessions (up to 100 per cent of wrong overgeneralizations in some sessions) and especially high error rates for some specific lexemes (up to 67 per cent of errors) as visible in the individual analyses of the longitudinal studies in Marcus (1995). In a dense corpus study, Maslen et al. (2004) found that, even though a large number of nouns are not overgeneralized, there is a substantial number for which periods of overgeneralization can be identified. Even though the overall mean of overgeneralizations is similar to the numbers found in Marcus (1995), Maslen et al. (2004) show convincingly that the overall mean is not the right measure to judge about development, especially without taking sample frequency into account. There is a clear development in the occurrence of overgeneralizing irregular forms in the high density corpus of one child ‘Brian’. The first overgeneralization occurs at age 2;02.00 with a mean of 5.9 per cent but then later on, overgeneralizations become more frequent up to 50 per cent at age 2;08.27 (Maslen et al. 2004: 1323f). Results of this study clearly indicate that there is a strong negative correlation between overgeneralizations and frequencies in the input. The nouns with the lowest input frequencies show higher overgeneralization whereas the most frequent irregular nouns are not overgeneralized at all. In a comparative study of English and German plural formation Köpcke (1998) suggests that the learning process can be better described by schema-learning than

inflectional morphology in language acquisition

361

rule-based learning. The German plural is very complex compared to English plural formation. The German plural has no clear default, but rather eight different classes of plural formation with phonologically unrelated alternants (Behrens 2002). Behrens supports the analysis of Köpcke in showing that overgeneralizations vary strongly across different plural classes, some are error free and others show overgeneralization rates of up to 40 per cent. Snow et al. (1980) in a study with older Dutch children aged 7 and 12 years revealed that plural allomorphy in Dutch, which lies somewhere between German and English in complexity, is acquired in a piecemeal fashion. Their results suggest that children learn plural marking by becoming familiar with exemplars. This familiarity then leads them to derive generalizations across lexemes, which one might call rules. Thus, according to Snow et al., vocabulary acquisition helps in learning morphological rules. The question then arises how children behave in languages with even more morphological irregularity than German. Blount (1988) conducted a cross-sectional study with twenty-two 3–14 year old learners of Luo (Nilotic, Kenya). Luo has two plural suffixes -e or -ini, the former is the one that occurs most frequently and is added to word final C and replaces final V if existent (Blount 1988: 226). In addition to the plural suffix -e, complex stem changes occur which Blount summarizes by two operations (i) ‘change in voicing or mode of articulation on the word final element’ and (ii) ‘addition of homorganic C’ (1988: 227). In principle what happens is that for every stem-final consonant the child has to remember complex stem changes. Blount finds that in addition to these complex but regular processes the child has to cope with a large number of irregular forms, based on suppletion, irregular stem change or deletion of final V, etc. To test how children cope with such a complex system, Blount conducted a test with nouns and nonce nouns. At age five in only 57 per cent of the cases children chose the correct form as used in the adult language. Children from age 6 were correct in 85 per cent of the lexical items and starting from age 10 were correct in 100 per cent of the cases. For the nonce nouns even at age 8 children were correct in only 57 per cent of the cases and the 14-year-olds achieved only 86 per cent. The adults behaved as expected with 100 per cent of correct forms. Blount identified morphophonological processes as the relevant processes for plural acquisition and children had varying difficulties with different noun types. Different morphophonological processes pose different difficulties to the children, stem alternation in nonce forms in particular was not produced correctly in more than 64 per cent of the cases, even in 12-year-olds. There was a significant increase of correct answers with age and significant differences in the different morphophonological classes. Blount found a predictable order of acquistion of plurals in Luo. He first showed that children had more difficulty with nonce forms than with real lexical items of the same morphophonological build-up. Overall children showed greatest difficulties with stem alternations when stem-final liquids were involved such as muor → muoce (nonce word). These difficulties are paralleled in real lexical items such as wuor ‘shoe’ → wuoce (pl). The next most difficult morphophonological process

362

sabine stoll

was the voicing change of stem-final C as in okupu → okube (nonce), again paralled in real lexical items such as kitabu ‘book’ → kitapu (pl). Stem final nasals such as um → umbe (nonce) proved to be the easiest forms for all children of the tested age range. Results on relatively simple plural systems such as the English plural and also complex plural systems suggest that lexical and schema-based-learning can account better for the acquisition of plural systems than models that rely exclusively on default learning.

15.2.4.2 Noun classes Bantu languages are famous for their nominal classes. Most of the Bantu languages have several nominal classes and, depending on the linguistic analysis, plural and singular are either counted as two separate classes or as a single class. Nominal classes are of special interest for acquisition because singular and plural forms are marked equipollently with class-based allomorphy. Noun classes are generally marked by prefixes, which are morphologically easy to segment and they exhibit phonologically transparent agreement systems (Demuth and Ellis 2009). Still one could expect that children could approach this system by overgeneralizing one plural morpheme to other classes as hypothesized by Slobin (1985a). However, there seems to be no evidence for such overgeneralizations, as shown in Demuth (2003), who summarizes the literature on the acquisition of noun classes in seven Bantu languages. In fact children learning these languages exhibit very similar patterns of acquisition. Three-year-old children already seem to have no difficulties in applying the different noun classes in conversation and hardly any overgeneralizations are found. Thus, Demuth suggests that ‘singular and plural noun class prefixes are segmented as separate morphemes early on.’ (Demuth 2003: 211). She reports three stages of noun class acquisition moving from (a) ‘no prefixes’ to (b) ‘shadow vowel and nasal prefixes’ to (c) ‘full and phonologically appropriate noun class prefixes.’ (Demuth 2003: 211), and these stages seem to be the same in all the longitudinal data of the various languages, even though they are not necessarily sequential in occurrence. She exemplifies this process with examples from a 2;1-year-old Sotho child uttered on the same day. First the child said ponko and then later on in the day apoko. The adult form is li̸=ph´ɔqɔ ‘green corn stalk’ (Demuth 2003: 211). But as stated in Demuth (1989) and Ziesler and Demuth (1995), such stages are not meant as discrete units children have to go through but rather they can use all of them at different points of time. The question then arises why children omit prefixes inconsistently. One hypothesis is that children mirror the input. In Sotho and Basotho it was found that not only children but also adults speaking to children drop prefixes in the input to children in about 5 per cent of the cases. Overall they use many more prefixless nouns than prefixed nouns, amounting to 70 per cent prefixless and dropped prefix forms (Ziesler and Demuth 1995: 10, citing Connelly 1984 for the Basotho data). Thus, the majority of the Sotho input consists of prefixless nouns, which might have an influence on the early

inflectional morphology in language acquisition

363

prefix dropping of young Sotho learners. In a reanalysis of early noun class use in the input of two children at two recording times (Time 1 age 2;1–2;7, Time 2 age 2;7–3;2), Demuth and Ellis (2009) found that adults only drop prefixes very selectively, namely only in the context of coronal consonants when followed by some kind of agreement, and only in about 20–35 per cent of the possible cases. The two children in the study imitate this pattern from early on and produce coronal prefixes significantly less often than non-coronal prefixes. Interestingly, the two children showed different overgeneralization patterns. One child dropped coronal prefixes regardless of the presence of agreement at the first recording point and then at recording Time 2 the child dropped coronal prefixes only when agreement was present. The other child showed a different pattern, at Time 1 she dropped prefixes on coronals and non-coronals that contained agreement and at Time 2 she dropped them mainly with coronals when agreement was present (Demuth and Ellis 2009). Thus, the two children showed different pathways of acquisition and this clearly demonstrates that individual developments need to be analysed in detail before generalizations across children can be made. As Demuth and Ellis suggest, children make syntactic and phonological overgeneralizations, which shows that they do not simply copy their input.

15.2.4.3 Case The challenge for the language learner in acquiring case is to connect the formal expression of case as an inflectional category to the grammatical functions these formal markers express, such as subject or object. The tasks for children learning languages with different case systems such as ergative, accusative, or split marking differ considerably. We know most about the acquisition of case in languages that exhibit exclusively accusative alignment but we cannot be sure whether children learning different sytems, such as the ergative, use the same learning strategies. One of the most pressing questions in the study of argument structure acquisition has been whether children learn argument structure by relying directly on semantic notions such as agency as proposed by Pinker in his ‘semantic bootstrapping hypothesis’ Pinker (1984), or whether they rather imitate the morphological patterns of their surrounding speech. Agency in the bootstrapping approach is hypothesized to help the child in finding the subject of a sentence and then she will determine whether case marking is assigned by nominative/accusative or ergative/absolutive marking. Languages with ergative cases are a good test case for these two hypotheses. An innate agentivity bias would be supported if children learning a language with split ergative alignment would generalize the ergative to all contexts with an agent. This would mean that they generalize according to semantic criteria. Studies on a number of non-related languages with ergative case marking, such as Hindi (Indo-European, Narasimhan 2005, Samoan (Austronesian, Samoa, Ochs 1982), Kaluli (Papua-New Guinea, Schieffelin 1985), Basque (Ezeizabarrena and Larranaga 1996) and a large number of other languages (see Bavin and Stoll 2013) do not find such semantically motivated overgeneralization errors, even though other kinds of

364

sabine stoll

overgeneralization errors, based for example on word order patterns such as in Kaluli, do occur (Schieffelin 1985). Hindi, for instance, has split ergativity conditioned by aspect. Only agents in perfective transitive sentences receive ergative marking, all other agents are marked with the nominative. If children relied on a semantic notion of agency one would expect overextension of the ergative case to imperfective transitive actions or agentive intransitives. Such overextension was not found by Narasimhan (2005) who studied the use of case in longitudinal studies (1;7–3;9) of three Hindi children. In languages with both accusative and ergative case marking, there is ubiquitous variation in case. In many languages assignment of cases depends on lexical valency classes and accordingly there is abundant variation in how fast and in what way case is learned. In the following paragraphs, the focus is on three languages which are morphologically accusative but mark case in different positions and ways. In German, case is predominantly marked on the determiner. Determiners are perceptually not salient, they cliticize and are therefore blurred in speech occuring in several different contextual realizations. An additional feature which might complicate the learning of such a system is the strong degree of syncretism of forms. These features might predict a later acquisition. The findings on German case acquisition vary. Eisenbeiss et al. (2005) in studying eleven recordings of spontaneous speech of five children (age 2;6–3;6) found a difference in the error rates of what she calls structural case, that is, default case markings of grammatical functions and lexical cases. Errors with structural cases were rare but errors with lexical cases—cases assigned by individual verbs—were ubiquitous (Eisenbeiss et al. 2005: 28). From this they conclude that children from early on are sensitive to syntactic constraints. Wittek and Tomasello (2005), in a nonce word experiment with twenty-four German children (mean age 2;8, range 2;7–3;0), found that half of the children were productive with the nominative case and two-thirds with the accusative case. However, at the same time, only one-third were productive with marking the active transitive and only one-fifth of the children were productive marking passive constructions, which shows that there is quite some difference in the understanding of case in different constructions. They conclude that in the study of case marking, different functions need to be analysed separately as indicated by the difficulties children have with the role of case for some sentence-level syntactic functions. Further, children in these age groups were far less productive with the dative case. Wittek and Tomasello (2005) conclude that children of this age have only partly mastered the functions of case and they first concentrate on local cues such as case marking within an NP before abstracting to sentence-level constructions. Szagun (2004) in a longitudinal study of six German children between 1;4 and 3;8 found that perceptual and frequency factors interact in case acquisition. Not all cases are learned at an equal pace: the nominative has the fewest error rates, followed by the accusative and then the dative. This goes along with the frequency of use of these cases in the input. There are a large number of substitutions of masculine definite

inflectional morphology in language acquisition

365

article den (acc)/dem (dat) with an overall mean of substitutions of 83.6 per cent (Szagun 2004: table 6). There is a further dissociation in the nominative and accusative masculine between the definite der/den and the indefinite paradigm ein/einen, with more errors in the indefinite paradigm (Szagun 2004: 26). The findings of the Wittek and Tomasello (2005) and Szagun (2004) suggest that in German the development of the case system is not conditioned by syntactic categorization and rule-based learning but rather a number of interacting factors such as formal realizations of case, acoustic properties, saliency of the markers, and frequency in the adult language are relevant for the piecemeal acquisition process. An important additional factor seems to be the ambiguity of forms presented in the input. A study by Pelham (2011) of the input to German and English children suggests that the ambiguity in the forms presented in the input varies significantly in German and English and this might influence the error rates in acquisition. Pelham, in studying the input of twenty-four German and twenty-four English children in corpora, finds ‘that case-ambiguous pronouns averaged 63.3 per cent in English, compared with 7.6 per cent in German.’ (Pelham 2011: 1). In the input of the English mothers the case ambiguous pronouns you and it were the most frequently used pronouns (approximately 60 per cent). The case ambiguous pronouns in German are es, ‘it’ and sie ‘she’ (nom and acc) and a few others but they make only about 7 per cent of all the pronouns uttered by the caretaker. However, in German, case ambiguities of determiners such as das (nom/acc.sg.n) or die (nom/acc.sg.f) are the most frequent determiners (77 per cent) in the input. Both English and German children make case errors in the case marking of pronouns and German children also do so in determiners (see Szagun 2004). However, how these frequency distributions relate to the development in children needs to be determined in the future. In Polish, case is marked on all components of the noun phrase. Modifying adjectives, demonstratives, and possessives agree with the noun. The system is fairly elaborate with seven cases in the singular and plural. The forms of the cases interact with inflectional class and partly (in the accusative masculine and the genitive) with semantic factors. Dabrowska ˛ (2001) in a longitudinal study on the acquisition of the Polish genitive has shown that inflectional subsystems can differ in how they are learned and default learning is ‘attributable to the properties of the relevant inflectional subsystems, not to the predispositions that children bring to the language learning ´ task.’ (Dabrowska ˛ 2001: 545). Dabrowska ˛ and Szczerbinski (2006) measured phonological diversity of the individual case affixes and the frequency of these forms in the input. They found that there is a difference in the various endings in regularity, overall frequency, and phonological structure. Masculine inflections showed more diversity than feminines or neuters. Token frequency was the best predictor of productivity in the younger children (2- and 3-year-olds), while later on in development phonological diversity was more important. They further found a strong effect of lexicality, but ´ no dissociation between regulars and irregulars (Dabrowska ˛ and Szczerbinski 2006: 595). Krajewski et al. (2012: 30), in a dense corpus study of one Polish child, support the findings of other studies of low overgeneralization rates. Further, they show that

366

sabine stoll

the Polish 2-year-old of their study can use all thirteen case–number combinations, although the range of nouns used is not the same as found in the surrounding adult speech. In addition, the contexts in which the child uses the different case markings are more restricted in the child than in the adult data. Thus, even though all the categories are found in early acquisition, they are restricted to specific lexical items, which supports early item-based learning, and shows that children’s proficiency is not the same as those of adults. These studies on the acquisition of Polish case show the importance of individual case studies to disentangle variables that are relevant at different times of acquisition. Unlike English or German, in Japanese case is marked on the head by particles. In an imitation and production study with children between 3;10 and 6;8 Hakuta (1982) found differential mastering of inflections with respect to the position in the sentence. Hakuta concludes that in Japanese acquisition word order strongly interacts with the acquisition of morphology. The studies presented here show that in typologically different languages several language-specific factors influence acquisition. The main important factors for the acquisition of case are homonymy, position in the sentence, phonological properties, and frequency patterns. Language specific combinations of these factors influence the acquisition of case in the individual languages, whereas agency seems to play only a minor role.

15.2.5 The Acquisition of Some Verbal Categories 15.2.5.1 Tense and aspect Tense locates the time an event takes place with respect to the time of speaking. Aspect in contrast to tense focuses on the temporal organization of the event itself, for example whether it has been completed or whether it is continuing. It is very common in the languages of the world that one morphological form expresses both tense and aspect. However, not all languages express tense or aspect (Dahl and Velupillai 2005). Most commonly, if tense is expressed, the past tense is overtly marked in contrast to nonpast forms. In the following the focus is first on the acquisition of tense and then on the relation between tense and aspect. As for most other grammatical categories, the majority of acquisition studies focus on English, even though English has rather limited verbal morphology. English has a clear default in past tense formation and a rather limited number of verbs that are irregular, which however are very frequent (see Bybee and Slobin 1982). Marcus et al. (1992) explain the acquisition of this system by a dual mechanism account which consists of a general rule for regular verbs and a number of irregular verbs which are learned as complete word forms by rote. As soon as a child encounters an irregular form according to their theory, rule application of the default is blocked (blocking hypothesis) and the child remembers the irregular form even after a single encounter. However, it is well-known that children often use both forms of the

inflectional morphology in language acquisition

367

same verb, the correct irregular form and an overgeneralization (e.g. eated and ate), sometimes for extended periods of time. To account for this fact Marcus et al. (1992) admit periodic failure of the blocking account due to memory-retrieval issues (see Maratsos 2000: 185 for a detailed discussion). However, they predict that overgeneralizations are rare, whereby rare use is defined in a rather broad sense, namely if less than 50 per cent of the irregular verbs are overgeneralized, they consider their use to be rare (Marcus et al. 1992: 18, 34). In analysing a mixed corpus of ten longitudinal studies and fifteen cross-sectional studies with mostly single samples with very few verbs, Marcus et al. (1992) claim that overgeneralizations of irregular foms are indeed rare. The mean of the twenty-five children (including both longitudinal and crosssectional data) tested levels to 4.2 per cent but there is an extreme range of rates from 0 per cent of overgeneralizations in those children with only a few verbs in their crosssectional data up to 24 per cent in one child (Abe) from a longitudinal study with the biggest numbers of verbs in his corpus. The inclusion of sampling snapshots (of the fifteen cross-sectional children) with children uttering only a few verbs is problematic in such an analysis and leads to an averaging artefact. This is confirmed by much higher rates of overgeneralization in the longitudinal data, where the highest rates of overgeneralization in a given month amount to 47.6 per cent (Abe). Maratsos (2000) tested the blocking account hypothesis and the claim of low overgeneralization rates in English past tense acquisition. He shows in an analysis of the longitudinal data of three English learning children that overgeneralization is not a rare phenomenon at all. One of the three children showed a lot of overgeneralizations and was very consistent in his behaviour, whereas the other children restricted their overregularizations to specific periods as in the data of Marcus et al. (1992). The overall overgeneralization rate was 24 per cent out of 1942 past tense tokens taken from 105 hours of recordings (Maratsos 2000: 191). Additional support of the finding that overgeneralizations are not at all rare was given in a high density study by Maslen et al. (2004). Maslen et al. found that overgeneralizations vary in frequency ranging from 0 per cent to 43 per cent in the thirty-four consecutive samples of their data. A reanalysis of Marcus et al. (1992) data by Stemberger (1993) suggests that acquisition of past tense in English can be explained by a single mechanism account. He shows that overgeneralizations rely on phonological features. The likelihood of overgeneralizations depends on what he calls vowel dominance, that is, if the base vowel is dominant (e.g. run–ran) regularization is likely, whereas if the past tense vowel is dominant (e.g. come–came), overregularizations are unlikely. Vowel dominance was determined by a priming task inducing adult slips of the tongue, which showed clear predictions of vowel overgeneralizations. This shows that irregulars are not treated as whole verb forms as assumed in Marcus et al. (1992). Similar results are found by Marchman (1997) in a production task on the acquisition of the English past tense with seventy-four children ranging from 3;8 to 13;5. Marchman found that phonological characteristics play a role both for the formation of regular and irregular patterns. Independent support of these findings was given by Plunkett and Marchman (1993) with a trained network which produced both early

368

sabine stoll

correct output and then subsequent regularized forms with a single mechanism strategy. These studies provide counterevidence to the assumption of dual mechanism approaches claiming that overgeneralizations are rare and that irregular forms are learned as unanalysed wholes, in other words it is not the case that children rely on a ‘phonologically-insensitive default mechanism’ (Marchman 1997: 283). An interesting result of Maratsos’ (2000) study that deserves especial emphasis is the strong individual variation found in the use of overgeneralizations also nicely illustrated in the analysis of the longitudinal data in Marcus et al. (1992). This suggests that there is much more variation in the pathways of acquisition than assumed by dual-mechanism approaches. Further evidence of overgeneralizations across the board comes from a study on the acquisition of Italian verb morphology. Italian children, who have to learn three conjugation classes make errors both in productive and non-productive classes, and no categorical difference is found between these errors (Orsolini et al. 1998). Interestingly, in this study Orsolini et al. found that older children also make such overextension errors. The study included two experiments, one narrative experiment with 100 children aged 4–10, and a sentence completion task with sixty children aged 5, 6, and 8. Similar results were obtained in a priming experiment and an elicitation task using novel verbs (Orsolini and Marslen-Wilson 1997). Additional support for a usage-based, single-mechanism account comes from Norwegian and Icelandic (Ragnarsdóttir et al. 1999). As stated by Ragnasdóttir et al. (1999: 584), in both Norwegian and Icelandic most past tenses are regularly built by suffixation and there is a small number of irregular verbs in both languages (approx. 4 per cent). However, there are two types of weak verbs and the distribution of these two classes differs in Norwegian and Icelandic. In both languages there is one predominant class (56 per cent in Norwegian and 75 per cent in Icelandic) but in Icelandic the distribution is more pronounced. In general, Norwegian morphology is less complex than Icelandic morphology which inflects for a large number of verbal categories and, unlike Norwegian, it also includes vowel change in weak verbs. Ragnasdóttir et al. in a cross-sectional study of 4–6- and 8-year olds suggest that morphological complexity, especially in the very early phases of acquisition, is an important feature in the acquisition of the past tense in the two languages. At age 4 the Icelandic children’s performance lagged behind the Norwegian children’s performance. At age 6, the performance was similar. However, they did not find evidence that either of the two weak classes was used as a default. The most important factor in predicting development in these two languages were type and token frequency and phonological regularities (Ragnasdóttir et al. 1999: 609). In a language like German a further issue in past tense formation arises, which provides difficulties for children well into their preschool years. In the German present perfect children have to choose between two auxiliaries. An experimental study with nonce verbs by Wittek and Tomasello (2002) has provided evidence that at around 2;6 children used past participles correctly but children up to age 3;6 have difficulties in choosing the correct auxiliary.

inflectional morphology in language acquisition

369

In Turkish past tense formation, yet another complication comes into play. Past tense markers are intertwined with evidentiality. The marker -DI is used if the speaker had direct experience of the event and the marker -mI¸s if the speaker has had no direct experience. This is a semantically rather complex category and every time a child uses such a form s/he needs to decide which of the two situtations holds true. One can thus expect that this is a category which is learned rather late in development. Indeed children first used the past tense marker -DI when focusing on direct evidence AksuKoç (1988). Only after 3;8 did children become adult-like in their use of these forms. Results of these studies on typologically unrelated languages show that the acquisition of tense is rather a slow and piecemeal process rather than a rule-based, general process. Cross-linguistic research on tense–aspect acquisition led to a major debate as to whether tense or aspect are acquired first. It was advocated by some researchers that the early tense forms children use in fact denote aspectual meanings only. The strong version of this ‘aspect hypothesis’ was disconfirmed (Shirai and Anderson 1995). Nevertheless, a strong interaction of tense and aspect and/or Aktionsarten (i.e. lexical aspect) was supported cross-linguistically (Antinucci and Miller 1976; Weist et al. 1984; Weist 1986). A strong correlation has been detected in the use of perfective aspect, the past tense, and telicity. This pattern has been found in English, French, Italian, Greek, Japanese, Mandarin, Polish, Russian, and Turkish and with a range of different methods: longitudinal studies and a number of diverse experimental comprehension and production studies have confirmed this pattern (for a summary of this research see Li and Shirai 2000). Several researchers have hypothesized that the distributions of tense/aspect forms found in the speech of children actually mirror the distributions in the input (Brown 1973; Stephany 1981, 1985; Shirai and Anderson 1995). Stephany was the first to study the aspectual distributions in the input and found a strong correlation between tense and aspect, that is, past with perfective and telic verbs vs. present with imperfective and atelic verbs in the adult data of her Greek corpus similar to the correlations in the data of the children. Shirai and Anderson (1995) hypothesized that this correlation characterizes adult speech in general. In a large-scale longitudinal study of aspectual use in Russian child-directed speech compared with child speech, Stoll and Gries (2009) found a strong patterning of the above mentioned correlations between aspect, tense and, Aktionsart in both children and their caretakers. However, the early correlations in child speech were much stronger than the correlations in child-directed speech. Only later in development, did the child adapt to the correlations in the input of perfective aspect with past tense and imperfective aspect with present tense. Overall correlations, however, might be too general since not all verbs are treated equally by the children. In experimental studies of the different Aktionsarten in Russian, it turned out that some Aktionsarten, such as for example the telic Aktionsart, are used and understood productively from early in development and others are tightly bound to very specific contexts and are acquired much later. Some less frequent and semantically more restricted Aktionsarten, such as the ingressive or delimitative Aktionsarten, are learned

370

sabine stoll

in a more piecemeal fashion, restricted first to specific limited contexts (up to at least age 6 but probably much later) and only later are abstract aspectual generalizations performed (Stoll 2005).

15.2.5.2 Passives Passives in early acquisition are of special interest because they focus on the patient of an action rather than on the actor, who is often semantically more salient. Thus passive constructions have been hypothesized to be cognitively more challenging for the young child than active constructions. Further, from a formal point of view, passives are interesting because in some languages not only is the morphology on the verb relevant to make an utterance passive but word order is also correlated with the presence of morphology. Thus, the task of using a passive seems to be more complex than forming an active utterance because of these interacting dimensions. In fact, passives have been claimed to be a later development in acquisition (Borer and Wexler 1987). Further, they are supposedly infrequent in the input, at least in English. Both longitudinal data and experimental data on English passive acquisition suggest that learning the different passive constructions is a slow process and that it takes many years before children reach adult competence. Budwig (1990), analysing all passive constructions of the diary data of Bowerman, found passives as early as 2;2 but before age 3;6 relatively few passives (twelve in total) were used. The two children of the diary study seem to distinguish be- and get-passives but restrict the be-passive to an activity scene with an unknown agent (Budwig 1990: 1247). Further, it has been shown experimentally that children produce passives early but with restrictions (Maratsos et al. 1985; Marchman et al. 1991). Both Maratsos et al. and Marchman et al. found that children prefer the passive with prototypical transitive action verbs. Marchman et al. provide evidence that, on the one hand, children often use an avoidance strategy in choosing an alternative construction and, on the other hand, if they do use a passive they avoid the full passive, that is, they use a truncated construction without the ‘by-phrase’ significantly more often than adults do. Thus, the acquisition of the passive in English is a process that goes on to late childhood (Marchman et al. 1991: 89). The same is commonly assumed for German and Hebrew which have been commonly cited in the literature in exhibiting late acquisition of the passive (Mills 1985 and for Hebrew see Berman 1985 citing a pilot study by Ariel). However, both of these studies are based on extremely limited datasets and newer data do not seem to corroborate these claims. More recent research by Abbot-Smith and Behrens (2006) using a high density longitudinal corpus of one child has shown that these claims do not hold for the German passive, which is learned much earlier than assumed in the literature. The German seinpassive (similar to the English adjectival passive) is acquired by the child around age 2;2 and the werden-passive around age 2;7. Results by Abbot-Smith and Behrens illustrate convincingly that we need larger datasets since small datasets, as used for example in Mills (1985), can bias the results in not capturing low frequency constructions including passives.

inflectional morphology in language acquisition

371

The assumption that passives universally are acquired late has also been challenged by a number of studies on typologically unrelated languages. Children learning Sesotho (Bantu, Lesotho) are productive with the passive as early as 2;2 (Demuth 1989, 1990). Passives are formed by NP movement to subject position and the addition of a passive morpheme -o- or -uo- to the verb. Thus, Sesotho passives are equally complex as English passives. Similar to English the by-phrase can be expressed optionally. This is done with the marker -ke. Two important differences to English, however, are that the passive in Sesotho is very frequent in the input and second that there is a typological difference in the function of passives. In Sesotho only old or given information can be in subject position and subjects cannot be questioned. The only way to ask about transitive agents is through passives, which remove the agent from the subject position. Thus, passives are a structure that needs to be learned early to enable the child to ask wh-questions. In fact by age 2;2 children do use and understand full passives (Demuth 1989, 1990). Similar results were found in a longitudinal study on Zulu acquisition of ‘several children between 1;10 and 3;6’ (Suzman 1985). In Zulu the passive is also used pervasively, and similar to Sesotho it is used for questioning subjects in Wh-questions. Further, recipients and benefactives can be passivized, and passivization is used to express object relatives with full noun subjects, and stative adjectives. There are also verbs with active meaning and passive form and a number of other non-canonical passive constructions (Suzman 1985: 132). Suzman found the first instances shortly after age 2 in fixed phrases. Children from age 2;6 onward use some passive forms spontaneously but only in restricted contexts, namely if something happened to them personally or to somebody they knew. Usually, it was used when somebody was affected negatively. Thus the construction is restricted to situations which are very salient for the child. This contextually retricted use is then extended to other situations later on in development. Children learning Inuktitut use passives from early on as well and with much greater frequency than, for instance, English children do (Allen and Crago 1996). Similar to English in passive constructions, the word order changes, that is, the object moves into subject position, and there is a passive affix on the verb. The adjectival passive equivalent is formed for a subset of verbs by adding the perfective affix -sima (Allen and Crago 1996: 133). Allen and Crago (1996) in a longitudinal study of four children (age 2;0–3;6) found that both short and full forms and passives with action and experiential verb roots are used productively (in the sense of generalizing the forms to new items not in the sense of showing adult-like behaviour) by children from 2;0 onward. Evidence for productivity is, however, difficult to maintain given the small numbers of full passives uttered by the children. Allen and Crago provide further evidence for productivity with passives, namely they found three passives with innovative forms, that is, forms the children could not have possibly heard from adults, such as for example the use of a passive form on a noun without first incorporating it into a verb. The three examples were all ungrammatical and would not occur in that form in the speech of an adult.

372

sabine stoll

Similarly, in K’iche Mayan, in which passive constructions are reported to be frequent in the input, children use passives early. Pye and Quixtan Poz (1988) found that children from age 2 onward use passives ubiquitously, much more often than their English peers. These results on a variety of typologically unrelated languages strongly suggest that the passive construction is cognitively no more complex than the active construction but instead language specific factors such as frequency in the input and range of application are important factors or even determiners for the onset of this construction.

15.3 Summary

.............................................................................................................................................................................

Studies on the acquisition of a large variety of morphological categories in typologically different languages support children’s great sensitivity to (i) input frequency, (ii) functional needs, (iii) context of occurrence and use of morphological markers in specific constructions. Irrespective of the language they learn, from early on they seem to imitate the prevalent patterns of their surrounding speech and rely on smallscale generalizations that occur both in the productive patterns and in irregular forms. These generalizations strongly interrelate with the grammar and frequency patterns of individual languages. The approaches children take not only vary across languages but also within languages. The research summarized in this chapter does not provide evidence for general rulebased learning, but rather supports an approach that builds on small-scale schemabased acquisition strategies that are strongly connected with distributional learning mechanisms (e.g. Tomasello 1992, 2003; Lieven et al. 1997). Independent support for the importance of distributional learning comes from a recent modelling study of early comprehension and production based on the input children receive in a number of typologically different languages. McCauley and Christiansen (2011) show that both comprehension and production can be better explained by the use of stored lexical chunks than by abstract rules operating over word classes. One of the main challenges of current acquisition research is to find out whether comprehension and production strategies in real life can be explained by similar mechanisms or whether they are indeed different procedures. Cross-linguistic research shows that there is strong variation both in the speed a linguistic feature is learned and in the learning strategies chosen. This suggests that there are many variables involved that need to be included in explanations of morphology learning. One component that plays centre stage is meaning as suggested by Naigles (2002). The assignment of meaning to a specific feature is much more complex than detecting repetitive structural patterns based on distributional regularities in the input. Comprehension studies with prelinguistic children focusing on pattern detection have

inflectional morphology in language acquisition

373

found that children are extremely good statistical learners and from early on infants work with immediate generalizations (Saffran et al. 1996; Saffran 2002). Saffran et al. found that eight-month-old children are able to segment words from speech relying exclusively on ‘statistical relationships between neighboring speech sounds’ (Saffran et al. 1996: 1926). These results are in contrast with most production studies, in which children are often found to restrict their use of grammatical morphemes to specific lexical items or constructions before generalizing so similar contexts (e.g. Tomasello 1992; Lieven et al. 1997). Naigles (2002) proposed that the different results are due to the fact that in these early processing studies with prelinguistic infants no meaning is involved and meaning has many more dimensions than the meaningless input consisting of repetitive patterns which is tailored to a set of controlled variables. An additional factor that has been neglected so far concerns the different distributional patterns of linguistic features provided in comprehension experiments focusing on pattern detection, and the real-world contexts which are the basis for generalizations of meaning necessary for production. In comprehension experiments children make immediate generalizations based on the data provided. The exposure to the test items is specifically tailored to the question addressed and the items are usually presented in a short timeframe. However, in free conversation, which children are exposed to in their daily lives, repetitions of forms are rather unpredictable and susceptible to contextual restrictions. Sometimes there are long intervals between the occurrence of the same or similar form–meaning pairings. This might complicate generalizations in natural language learning or at least delay them. We do not yet know enough about the role of these distributions and the impact of memory issues with respect to generalizations. Thus, it is still unclear whether in production children use the same strategies as in early comprehension but due to the natural language context these strategies are difficult to detect or, instead, whether they use other strategies such as item-specific learning strategies which make selectional use of distributions before generalizing. This is an issue for future research. Cross-linguistic research in studying the acquisition of inflectional morphology has made significant progress in the last decades. However, there are two areas in which further advances are very much needed. First, we need to learn more about under-researched areas of morphology acquisition. There are many areas of morphology for which acquisition is unknown. Second, research is still very much concerned with the acquisition of language-specific categories rather than taking a more comparative stance. To learn more about different strategies in typologically different languages, further comparative studies—both experimental and naturalistic—would have an impact on the detection of strategies applied in typologically diverse languages. With experiments we can test productivity in the sense of whether children are able to make systematic generalizations independent of lexical items. From longitudinal data we can learn about children’s proficiency, that is, we can find out from their actual behaviour when they behave in an adult-like way. Thus, proficiency describes the learning process as it occurs. To measure proficiency, the field has been developing new methods of analysis such as the entropy measure mentioned in Section 15.2.1. The discussion of both productivity and proficiency is

374

sabine stoll

crucial because even if we know that a child is able in principle to produce a specific form in an experimental context, but in naturalistic situations never actually does so, there are two possible interpretations. The first is that our sample was too small and we did not detect the occurrence. The second, however, is that even though the child is able in principle to produce a given form, in reality she does not yet do so spontaneously. If the latter alternative holds, the question arises why this is the case and in turn asks for the development of new methods and designs to answer this question. This is one of the major tasks of future research.

chapter 16 ........................................................................................................

DISORDERS ........................................................................................................

matthew walenski

16.1 Introduction

.............................................................................................................................................................................

The study of the neurocognitive basis of inflection involves a number of important issues. First, what computations does inflection depend on (e.g. in the transformation from a stem to an inflected form)? How and when are these computations carried out in real time? Which cognitive systems perform the computations, and what are their biological bases (e.g. anatomically defined structures, neurotransmitters, genetic factors, etc.)? Patients with neurocognitive disorders provide an important source of evidence by which to examine these issues. For each issue, we can ask whether there might be similarities in the language profiles and inflectional deficits (if any) across disorders that have similar pathologies, as well as whether there might be different performance patterns for different aspects of inflection. For example, are there differences between production and comprehension? Do nouns differ from verbs? Are these differences language specific, or do language universal patterns emerge? Are the computation, processing, and biocognition of inflection distinct from those of other aspects of language (e.g. phonology, syntax), or do common mechanisms cut across these domains? Likewise, are the biological bases of inflection specific to language, or are they domain-general, also subserving non-linguistic cognitive functions? If they are domain-general, which other cognitive functions relate to inflection? This chapter reviews research on these issues in light of the well-studied contrast between regular and irregular forms. In order to ease comparisons across disorders and languages, I focus narrowly on studies of inflected verb forms using elicited production tasks.

376

matthew walenski

16.2 Defining Regularity and Irregularity

.............................................................................................................................................................................

In order to examine the regular–irregular distinction from a neurocognitive perspective, clear criteria for identifying regular and irregular forms cross-linguistically need to be established. General definitions for each category are provided in this section, with specific details in Tables 16.1–16.5 for each of the languages discussed in this chapter. A regular inflected verb form is one that follows the default pattern for the language (Clahsen 1999; Pinker and Ullman 2002a). The default pattern is the most readily generalizable pattern (e.g. adding -ed to a verb stem for English past tense), for example to new words entering the language (fax → faxed) or to novel forms that do not have any meaning or prior representation in memory (blick → blicked). In many languages, including English, the default stem pattern is also the pattern that applies to the greatest number of verbs in the language, although this is not necessarily the case in every language (Clahsen 1999; Pinker and Ullman 2002a). Irregular inflected verb forms are those which deviate from the default pattern (Clahsen 1999; Pinker and Ullman 2002a), at least for the languages discussed here, if not for all languages. Irregular forms can involve stem allomorphy (e.g. bring → brought), sometimes to the point of full suppletion (go → went). In other cases, irregulars add non-default affixes (e.g. the German past participle geruf-en ‘called’ adds the non-default affix -n instead of the default -t). Semi-regular verb forms (also referred to as pseudo-regular and mixed forms) may be distinguished as a category in their own right. These forms are regular–irregular hybrids that involve stem allomorphy plus addition of the default stem-forming affix (e.g. German, Italian). Semi-regular forms have been less well studied than either regulars or irregulars, and are often grouped together with either fully regular or fully irregular items in experimental work. Note that languages differ in whether they obligatorily add agreement affixes (e.g. for person, number, and/or gender agreement) to particular verb forms. For example, Italian past participles add gender/number agreement affixes, whereas German and English past participles do not. These affixes do not factor into whether a given form in a particular language is classified as regular, irregular, or semi-regular, which is based solely on stem formation.

16.2.1 Theories of the Neurocognitive Basis of Inflection Multiple theories of language and cognition have offered explanations for the neurocognitive basis of inflection, broadly, and the regular–irregular distinction specifically. Here I present brief overviews of two perspectives that each make explicit, specific predictions regarding the relationship between inflection and its neurocognitive basis, but nevertheless provide quite distinct perspectives on this issue. Figure 16.1 provides a glimpse of relevant anatomical landmarks.

disorders (a)

(b)

(c)

(d)

377

fig. 16.1 Regions and structures of the brain: (a) Cortical regions, including gyri and sulci. The frontal lobe lies anterior to the central sulcus, above the lateral sulcus. The temporal lobe lies below the lateral sulcus, going back to the occipital lobe at the back of the brain. The parietal lobe lies behind the central sulcus above the temporal lobe. (b) Brodmann’s areas, which correspond to areas with different cellular architecture (e.g. different cell types and densities). (c) The basalganglia, which consist of several interconnected structures, of which the caudate, putamen, and globus pallidus are indicated here. (d) The hippocampus, together with other structures it is closely connected to. Source: Walenski et al. (2005: 335).

16.2.1.1 Dual-system models Dual-system models come in many varieties, although all posit a principled distinction between the lexicon and the grammar (Baayen et al. 1997; Marslen-Wilson and Tyler 1998; Clahsen 1999; Pinker 1999). On dual-system approaches, the lexicon is viewed as a repository of the sound–meaning mappings of individual words, as well as all idiosyncratic linguistic knowledge and memorized linguistic forms; the grammar specifies rules by which smaller units combine sequentially and hierarchically into complex words, phrases, and sentences. One particular dual-system model, the declarative/procedural model (Ullman et al. 1997; Ullman 2004), makes specific predictions regarding links between biology,

378

matthew walenski

cognition, and language, and grew out of Pinker’s ‘Words and Rules’ model (Pinker 1999; Pinker and Ullman 2002a), itself an outgrowth of earlier lexicalist approaches to morphology (Jackendoff 1975). On this view, production of an inflected verb form begins with two processes in parallel: a search for the inflected form in the mental lexicon, and computation of the inflected form by grammatical rule. The mental lexicon is not just a rote list of memorized idiosyncratic forms, but rather depends on an associative memory that can generate novel forms based on similarity with forms already stored in memory (e.g. spling → splang from sing → sang, spring → sprang, ring → rang).1 Successful output from the lexicon (e.g. an irregular past tense like dug or splang) blocks application of the rule (Marcus et al. 1992; Pinker 1999). If the lexical search fails (e.g. for the regular verb walk or for blick or even spling) then the rule successfully outputs the regular form (e.g. walked, blicked, splinged). The rule is thus a default process, and applies whenever the associative memory system fails to produce an acceptable output. Not all regular forms are predicted to be computed by the grammar however. First, the past tenses of inconsistent regular verbs (i.e. verbs whose stems are phonologically similar to those of irregular verbs; e.g. glide → glided) appear to be stored, as do the regular past tenses of doublet verbs (dive → dived or dove) (Ullman 2001b; Pinker and Ullman 2002a). If this were not the case, lexically-based irregular output (dove; or a potential irregular form glid, by analogy to hid) would block rule-based regular production (dived, glided) (Pinker 1999; Ullman 2001b).2 Second, certain factors also influence memory-based storage of even consistent regular forms (i.e. regular verbs whose stems are not phonologically similar to those of real irregular verbs), including frequency and imageability and sex—that is, highly frequent or highly imageable regular forms show effects indicative of lexical storage, particularly for women (Alegre and Gordon 1999; Tabak et al. 2005; Prado and Ullman 2009). Third, patients with a grammatical deficit but preserved mental lexicon could in principle memorize regular forms as a form of compensation for their deficit, even though these forms may normally be computed grammatically. The declarative/procedural model holds that the mental lexicon and the mental grammar are linked to distinct neurocognitive systems: memorization of idiosyncratic lexical forms in the mental lexicon depends on the declarative memory system; the sequential and hierarchical structuring of smaller pieces of language into larger

1 Throughout the chapter I refer to ‘phonological patterns’ to describe subregularities among irregular forms. While such subregularities are captured by rules on some approaches (e.g. lexical redundancy rules; Jackendoff 1975), both of the theoretical approaches discussed in this chapter instead treat irregular subregularities in terms of analogical generalization in a pattern-associating memory (Pinker and Ullman 2002a). 2 Note that an individual speaker could be aware of only one form of any or all of the verbs that have doublet status in the language (i.e. as assessed over the entire language community). Thus while the regular past tense form for an individual doublet verb for an individual speaker might not depend on memory (e.g. if the speaker were unaware of the irregular form), this behaviour is predicted (and indeed, has been found) for statistically reliable samples of speakers and verbs (Pinker and Ullman 2002a).

disorders

379

units (in phonology, morphology, and syntax) by the mental grammar depends on the procedural memory system (Ullman et al. 1997; Ullman 2004). These two memory systems have different characteristics and depend on largely distinct neurobiological substrates (for a general overview of both systems, see Eichenbaum and Cohen 2001). Declarative memory underlies the (generally, but not necessarily) explicit learning and storage of knowledge about facts (semantic memory; e.g. George Washington was the first American president) and personally experienced events, encoded with respect to a particular contextual setting (episodic memory; e.g. I took the bus to work this morning). Learning new facts depends largely on the hippocampus and other medial temporal structures, whereas the long-term storage of facts depends largely on neocortical regions in the temporal and parietal lobes. Other brain structures also play a role in declarative memory, including portions of the basal-ganglia and Brodmann’s areas (BA) 45/47 in inferior frontal cortex, which underlie the selection and retrieval of declarative memories (Ullman 2006). Acetylcholine and oestrogen play important roles in the function of declarative memory (Ullman 2004). Procedural memory underlies the implicit learning and control of motor and related cognitive skills, particularly those involving sequences, such as riding a bicycle or playing the piano. This system is rooted particularly in circuits connecting the basal-ganglia with regions of cortex in the frontal lobe, although it also includes portions of inferior parietal cortex, superior temporal cortex, and the cerebellum (Ullman 2004). The activity of the basal-ganglia ultimately either inhibits or disinhibits cortical neurons and the functions that depend on those regions. Importantly, dysfunction in these circuits is expected to have similar consequences across domains, such that an inhibited motor profile (i.e. with activity reduced in amount or degree) should correspond with inhibited language processing, and a disinhibited motor profile (i.e. with activity increased in amount or degree) should correspond to disinhibited language processing (Ullman et al. 1997; Ullman 2004). Dopamine is a particularly important neurotransmitter for this system (Ullman 2004). In sum, with respect to language, this dual-system model predicts that the production of irregular and other forms that have an increased dependence on the mental lexicon will be particularly impaired in disorders affecting declarative memory (e.g. Alzheimer’s disease). In contrast, the production of regular forms, which have an increased dependence on the mental grammar, should be impaired in disorders affecting procedural/motor functions (e.g. Parkinson’s disease). In addition, the model is domain general, and predicts that damage to particular brain structures will affect language and non-language functions subserved by those brain structures. Thus impairments to memory-dependent morphological forms should co-occur with impairments to other lexical and declarative memory abilities, and impairments to rule-produced morphological forms should be found with other grammatical/sequencing deficits in language (e.g. in syntax or phonology), moreover corresponding to the specific type of motor/procedural memory dysfunction (i.e. inhibited or disinhibited).

380

matthew walenski

16.2.1.2 Single-mechanism models Single-mechanism models grew out of functionalist ‘usage-based’ accounts of language that place a premium on the roles of frequency and similarity in the description of linguistic systems (Bybee 2007), and assume no a priori distinction between lexicon and grammar. Instead, they posit that the learning and use of language depend solely on associative pattern matching over stored representations (Bybee 1985, 2007; Rumelhart and McClelland 1996; Seidenberg 1997; Joanisse and Seidenberg 1999; McClelland and Patterson 2002b). Connectionist theory offers a computational framework for the single mechanism view (Rumelhart and McClelland 1996; Seidenberg 1997). One connectionist single-mechanism account makes explicit predictions regarding dissociations between regular and irregular inflected forms and the cortical regions they depend on (Joanisse and Seidenberg 1999; Patterson et al. 2001; McClelland and Patterson, 2002b). On this view, the production of irregular past tenses depends particularly on word meanings embodied in temporal-lobe cortical regions (particularly the temporal pole), whereas the production of regulars and novels is posited instead to show a greater reliance on phonology and frontal-lobe cortical regions. It has been argued that the greater reliance of regulars on phonology follows from the claim that regulars are more phonologically complex than irregulars (Bird et al. 2003). A similar claim maintains that (at least some) regular morphological forms are phonologically irregular in English (e.g. beeped, which has a long vowel followed by a consonant cluster, violates phonological constraints on syllable size; whipped, with a short vowel followed by a cluster, does not), whereas (many but not all) irregular morphological forms are phonologically regular (e.g. kept does not exceed these limitations on syllable size, sold does) (Burzio 2002). For either reason, production of a regular inflected form should tend to place a greater burden on phonological processing than production of an irregular one. This single-mechanism account therefore predicts that temporal-lobe cortical damage leading to semantic deficits will impair irregulars, leaving consistent regular and novel past tense production relatively spared. A semantic deficit is also claimed to lead to deficits of strongly associated phonological patterns, thereby affecting inconsistent regulars and novel verbs that are phonologically similar to existing irregular verbs (Woollams et al. 2009). In contrast, damage to regions of frontal cortex is expected to result in phonological deficits (broadly) and deficits of regular and novel past tenses. In addition, deficient phonological processing is expected to impair strongly associated semantics (Thomas et al. 2001).

16.2.1.3 Summary The Declarative/Procedural (DP) model and the single-mechanism account discussed here both make explicit predictions for the processing of regular and irregular inflected forms by patients with focal brain damage. However, the two approaches differ greatly in the functional roles attributed to the specific brain regions and how these functions

disorders

381

relate to regular vs. irregular inflection (for additional review and discussion, see McClelland and Patterson 2002a, 2002b; Pinker and Ullman 2002a, 2002b). Both accounts link irregular forms to semantic memory, although precisely what is represented in semantic memory differs. According to the declarative/procedural model, sound–meaning correspondences are stored in memory for all words, and thus at least some form of word-level phonological representations are stored in memory for all words. These stored word-level patterns form the basis for the associative declarative memory system to generalize to new irregularized forms (e.g. for novel verbs, like splang, or potentially for inconsistent regulars, like squoze). In contrast, only word meanings are in semantic memory in the single-mechanism account, with each meaning linked to appropriate frontal-cortex-based phonological patterns. These links are the basis for this account to predict deficits at novel irregulars or inconsistent regulars in patients with semantic deficits. Despite these differences, similar predictions of deficits at irregulars and inconsistent regulars are made by the two models for patients with damage to temporal-lobe cortical regions. The models differ more in their predictions for regular inflected forms and patients with frontal/basal-ganglia damage. The declarative/procedural model claims that patients with frontal/basal-ganglia damage will have an impaired mental grammar affecting rule-governed aspects of language, in phonology, morphology (i.e. the default rules underlying regular affixed forms), and syntax. The single-mechanism account denies grammatical rules altogether, but predicts phonological deficits (broadly) for these patients, that impact regular morphological forms more than irregular ones (to the extent that the specific exemplars of each type in a particular experiment differ in phonological complexity).

16.2.2 Elicited Production Tasks In a typical elicited production task, participants are given one form of a verb with an incomplete sentence context, and instructed to fill in the blank with the appropriate form of the given verb. For example, the participant may be given ‘walk. Every day I walk over there. Just like every day, yesterday I ______’ and is expected to fill in the blank by producing walked. Accuracy is the most frequently reported dependent measure, although response time can also be measured. In these tasks, item- and participant-specific factors that may influence performance or affect interpretation of the results are controlled either by matching items and participants on the relevant factors—for example matching regular and irregular verbs on frequency and phonological structure (regular called and irregular told are matched on these factors) or by accounting for them when analysing the data, or both (for discussion, see Clahsen et al. 2003; Walenski et al. 2007; Prado and Ullman 2009). In one variant of the task (sometimes referred to as the Ullman task), patients read the stimuli silently to themselves and produce their response out loud (Ullman et al. 1997). In another variant sometimes referred to as the Center for Speech and

382

matthew walenski

Language (CSL) task, participants listen to the stimuli rather than reading them aloud (Tyler et al. 2002).

16.3 Disorders Affecting Inflection

.............................................................................................................................................................................

In order to cover a range of deficient performance found across disorders and languages with elicited production tasks, while revealing aspects of the patterns that remain consistent across disorders and languages, I discuss data in detail from nine disorders, grouped according to their biological profile and non-linguistic deficits. Six were chosen for having a reasonably well understood biological basis, with relatively focal brain damage, thus enabling a workable, clear view of the relationship between biology and language. The three remaining disorders have multiple cognitive deficits, and poorly understood etiologies in two cases, but nevertheless show patterns of deficient language performance consistent with those in the more circumscribed disorders. Elicited verb production has been investigated in all nine disorders in English, with additional data from Italian, Greek, French, and German.

16.3.1 Disorders Affecting the Temporal Lobe and Semantic Memory The three disorders discussed here all involve dysfunctional semantic memory and temporal lobe damage. According both to dual-system and single-mechanism models, patients with these disorders should show impaired production of irregular and other memory-dependent inflected forms, with spared production of default regular forms.

16.3.1.1 Alzheimer’s disease Alzheimer’s disease is a progressive neurological condition characterized by the presence of neuritic plaques and neurofibrillary tangles that spread throughout the brain as the disease progresses (Hebert et al. 2003; Levy and Chelune 2007). The tangles first appear in medial temporal structures, beginning in entorhinal cortex, and spread in early stages to the hippocampus and neocortical temporal lobe regions, while leaving basal-ganglia and frontal areas relatively spared (Arnold et al. 1991; Braak and Braak 1991; Kemper 1994). Episodic memory is the earliest and most severely affected function, although impairments in semantic memory are present from early stages of the disease in at least some patients (Hodges et al. 1990; Grossman et al. 1997; Hodges and Patterson 1997). Language output is anomic (patients have difficulty finding words) and semantically

disorders

383

impoverished, while articulation and syntax remain spared, at least until later stages of the disease (Kempler et al. 1987; Levy and Chelune 2007; Kempler and Goral 2008; Hoffmann et al. 2010). Elicited production of inflected verb forms has been examined in English and Italian. English (Table 16.1). Alzheimer’s patients with semantic memory impairments were impaired at the production of real and novel irregular verbs (dug, crove), but not consistent regulars or regularized novels (looked, plagged, crived) (Ullman et al. 1997; Cortese et al. 2006; Ullman, in press).3 In addition, patients with more severe semantic memory deficits had greater difficulty producing irregular forms than those with less severe memory deficits, but did not have greater difficulty with real and novel regularized responses (Ullman et al. 1997; Ullman, in press). Patients with Alzheimer’s also appeared to be impaired at the production of inconsistent regulars (squeeze) (Cortese et al. 2006).4 The errors that patients made also reflect damage to memory (Ullman et al. 1997; Cortese et al. 2006; Ullman, in press). Patients produced over-regularization errors (digged) to real irregular verbs at high rates, and patients with more severe semantic memory impairments produced more of these errors.5 Unmarked forms (i.e. where the form produced is uninflected, and is the same as the stem form given to the patient; dig) were also produced at high rates; virtually no over-irregularization errors (dag) were made. Italian (Table 16.2). For class II and class III Italian verbs, Alzheimer’s patients were impaired at the production of real (chiuso ‘shut’) and novel (schisso) masculine singular irregular past participles, while the production of real (ballato ‘danced’) and novel (carlato) regular class I masculine singular past participles was spared (Colombo 3 The results from Cortese et al. (2006) did not reach statistical significance, although they are consistent with a pattern of impaired production of irregular past tenses. Several factors may have contributed to the lack of significance. First, Cortese et al. included no-change irregulars (e.g. hurt → hurt), for which correct responses are indistinguishable from unmarked errors (and indeed, patients were reported to make unmarked errors on the task). Second, only a small number of items were included (seven consistent irregular items and nine inconsistent irregular items), giving the patients few opportunities for error in each condition. The net effect of these factors would be to work against the finding of an irregular deficit. 4 This finding is potentially compromised by problems with two of the nine inconsistent regular verbs (fit, heave). Heave is a doublet verb, with correct (but infrequent) hove in addition to correct heaved. Fit is a no-change irregular verb in American English (the study was conducted in Missouri), although the regularized form fitted (which was counted as correct) is the standard form in British English (which would render it a double verb, if the study participants knew the British form). As irregular and doublet verbs are independently expected to be impaired in Alzheimer’s disease, if their inclusion here is responsible for the observed effect, it might suggest that Alzheimer’s patients are not impaired on inconsistent regulars (contrary to the prediction that these forms depend on memory). However, removing these two items did not change the mean per cent correct score (re-calculated based on the item means provided in the Appendix to Cortese et al. 2006), suggesting that their inclusion may not have actually skewed the results. 5 Errors on inconsistent regular verbs consisted predominantly of ‘other errors’ which were not described for these verbs specifically, as well as a very small number of over-irregularizations (Cortese et al. 2006).

384

matthew walenski

Table 16.1 Past tense inflection in Englisha

Rule produced Consistent regular Denominal Novel Memory dependent Irregular Inconsistent regular Doublet Novel

Stem

Past tense

Characteristics of type

look fly plag crive

looked flied plagged crived

• stem + -ed suffix • default—applies when memory fails to produce an irregular form

dig freeze squeeze dive crive

dug froze squeezed dived/dove crove

• unpredictable changes to the stem, no suffix • real regular verbs in competition with real or potential irregulars forms are memorized to avoid blocking • novel verbs that are phonologically similar to existing irregulars may be ‘attracted’ to the irregular pattern

a The distinction between rule-produced and memory-dependent forms is based on predictions

made by dual-system models (see text). Importantly, not all regular forms are claimed to be rule produced. Regular past tenses are formed by adding the phonologically appropriate allomorph of the –ed suffix to the verb stem (Pinker 1999). Irregular verbs do not follow this default pattern, but are instead formed by various changes to the stem (Pinker 1999). The majority of verbs in English are regular, with only about 180 irregular verbs in the language (Pinker and Ullman 2002a). Consistent regular verbs are defined as regulars whose stems are phonologically distinct from the stems of irregular verbs (Ullman 1999). Inconsistent regulars are verbs that are phonologically similar to existing irregulars (e.g. squeeze is similar to freeze; squeezed is memorized to block production of squoze). Doublet verbs are verbs that have both a regular and an irregular past tense form. The regular form is memorized to prevent blocking by the irregular form. Denominal verbs are an example of systematic regularization, whereby expected irregular verbs (e.g. fly → flew) are instead produced with a regular past tense (Pinker and Ullman 2002a). Novel verbs generally follow the default regular pattern, but are sometimes irregularized when they are phonologically similar to irregular verbs (e.g. crive → crove follows the pattern established by dive → dove, drive → drove, etc.) (Prasada and Pinker 1993).

et al. 2009; Walenski et al. 2009). Production of semi-regular past participles (temuto ‘feared’) was also impaired, consistent with a dependence on memory for these forms, although the impairment was not as severe as for irregular class II past participles (Colombo et al. 2009). Performance in the production of irregular past participles in Italian correlated with performance in synonym generation, whereas neither regular nor semi-regular performance correlated with synonym generation (Colombo et al. 2009). The lack of a correlation for the semi-regulars is consistent with the suggestion that these forms may be partially rule-produced (Say and Clahsen 2002). Likewise, the production of real first person singular forms was impaired for irregular verbs from all three verb classes, as well as specifically for class II verbs with phonologically predictable augments (e.g. salgo ‘I climb’). Production of real regular

disorders

385

Table 16.2 Past participle and present tense inflection in Italiana

Regular Class I

Infinitive

Past participle Characteristics (masculine singular)

amare ‘love’

am-a-t-o

• fully default form—–root + thematic vowel + past participle affix + agreement affix • class I is default class, used for new words entering language

temu-t-o veni-t-o

• ‘within class’ regular form • root + class-specific (i.e. non-default) thematic vowel is memorized as an irregular stem • default past participle affix and agreement affix are added to the stem

Semi-regular Class II temere ‘fear’ Class III venire ‘come’

Irregular Class Ib Class II Class III

Regular Class I Class IIc Class IIId

fare ‘do’ fatt-o discútere ‘discuss’ discuss-o aprire ‘open’ apèrt-o

• past participle stem is changed unpredictably from stem of infinitive • no past participle affix is added • agreement affixes are added to stem

Infinitive

Present tense (first person singular)

Characteristics

amare ‘love’ temere ‘fear’ dormire ‘sleep’

am-o tem-o dorm-o

• stem + agreement affix • inflected form gives no information about class membership

tèn-g-o fin-isc-o

• stem + agreement affix • augment (between stem and affix) is predictable for class II but not class III verbs

andare ‘go’ potere ‘be able’

vad-o poss-o

udire ‘hear’

òd-o

• stem + agreement affix • stem is changed unpredictably from stem of infinitive

Irregular (augmented) Class II tenére ‘have’ Class III finire ‘end’ Irregular Class I Class II Class III

a Italian verbs are traditionally divided into three morphological classes, distinguished by the

thematic vowel: -a- (class I), -e- (class II), and -i- (class III) in the infinitival form of the verb. Class I is by far the largest class (containing the most verbs) and has almost no irregular members (Napoli and Vogel 1990). It is also the default class, used for new verbs entering the language (e.g. chattare ‘to chat on the internet’). In contrast, roughly 95% of class II verbs display some morphological irregularity (Say and Clahsen 2002). Class III has few irregular members, although the morphological patterns of class III verbs are more similar to those of class II than class I verbs (Vincent 1987; Napoli and Vogel 1990; Say and Clahsen 2002). (Continued)

386

matthew walenski

Table 16.2 continued Irregular verbs are defined here as those which undergo stem-changes from the stem of the infinitive to the stem of the inflected form. Regular verbs are defined as those that undergo no such stem changes. For the regular past-participle (e.g. the class I verb amare), the verb root (am-) is combined with the thematic vowel (-a- for class I), the past participle affix -t-, and a gender/number agreement affix (e.g. masculine singular -o), yielding the inflected form (amato). Most class I and class III verbs are regular in the past-participle, whereas most class II verbs have irregular (stem-changing) past-participles. Semi-regular verbs (referred to as pseudo-regular verbs in some publications) are defined as class II or class III verbs, which have a memorydependent stem (root + non-default thematic vowel; temu-, veni-), plus the regular -t- affix (Say and Clahsen 2002). Stems are never produced without an affix specifying gender/number agreement (-o). Regular present tenses add one of six person/number agreement affixes to the infinitival stem. Irregular present tenses are created by adding the (same) person/number agreement affixes to a stem that is different from the infinitival stem, with largely unpredictable changes. Some irregular class II verbs change their stems by adding a -g- augment before a person/number affix. This augment is added predictably (though this transformation is arbitrarily only predictable for class II verbs), to verbs whose infinitival stems end in the phonemes /l/, /ʎ/, /n/, or /ɲ/ (spelled -l-, -gl-, -n-, and -gn- respectively). Class III verbs may also add an augment (either -g- or -isc-), although it is not predictable whether a given class III verb will take an augment, nor which augment it takes (Vogel 1993). Effectively then, augmented class II and class III verbs are irregular. b Irregular class I past participles were not included in any experiments discussed here. c Regular class II forms were tested for novel verbs only (Walenski et al. 2009). d Regular class III present tense forms were not included in any experiments discussed here.

class I first person singular present tense forms was spared (Walenski et al. 2009). For novel verbs, Alzheimer’s patients were impaired at producing expected regularized present tense responses to verbs with class II infinitives (incròdere → incrodo) but not class I infinitives (carlàre → carlo) (Walenski et al. 2009). This surprising regular deficit may be related to a morphological class effect—the majority of irregular verbs in Italian belong to class II, and so indication of membership in these classes (e.g. by the thematic vowel in the verb infinitive presented to the participants) may lead to the expectation of an irregular form, hence a greater effort by the impaired memory system in Alzheimer’s patients to compute an irregular form, leading to a higher error rate (Colombo et al. 2004; Walenski et al. 2009). The errors in irregular forms were not predominantly regularization errors, unlike in English. Instead, patients made lexically based errors in Italian, such as substitution errors (i.e. producing the inflected form of a verb other than the one given; e.g. giùngere → arrivato ‘arrived’), distortion errors (i.e. phonological changes to the verb stem; e.g. giùngere → gunto), and irregularization errors (i.e. applying an incorrect irregular pattern to the given verb; e.g. giùngere → giutto) on real and novel irregular past participles and present tenses (Colombo et al. 2009; Walenski et al. 2009). Such lexical errors are consistent with the notion that the expectation of an irregular form leads to an increased dependence on an impaired memory, yielding erroneous output (Colombo et al. 2009; Walenski et al. 2009). Interestingly, this pattern of errors

disorders

387

was found particularly when the patients were given the infinitive of the verb (chiedere ‘to ask’), which identifies morphological class membership. When the patients were given first person singular present tense forms instead (chiedo ‘I ask’), which do not identify the verb’s morphological class, they produced a much higher proportion of regularized past participles with the default class I -ato affix (chiedato)—that is, similar to the English pattern (Colombo et al. 2009).

16.3.1.2 Semantic dementia Semantic dementia is a progressive neurological condition, currently recognized as one of three subtypes of fronto-temporal lobar degeneration, along with fronto-temporal dementia and progressive nonfluent aphasia (Neary et al. 1998). Neurologically, anterior cortical regions of the temporal lobes are particularly affected in Semantic dementia (Hodges et al. 1992; Patterson et al. 2006). The disorder manifests primarily as an impairment of conceptual knowledge with severe anomia, with generally spared phonology and syntax (Hodges et al. 1992; Grossman and Ash 2004; Tyler et al. 2004; Patterson et al. 2006). The production of inflected morphological forms has been examined in a few studies of English-speaking patients. English (Table 16.1). Semantic dementia patients were impaired in the production of forms dependent on memory—real irregular and inconsistent regular past tense forms—but were spared in the production of real and novel consistent regular past tense forms (Patterson et al. 2001; Tyler et al. 2004; Cortese et al. 2006; Patterson et al. 2006).6 The patients made over-regularization errors (digged) to irregular verbs at high rates, consistent with an impaired memory for irregular forms (Patterson et al. 2001; Tyler et al. 2004; Cortese et al. 2006; Patterson et al. 2006). Errors on inconsistent regulars consisted primarily of irregularized responses (e.g. squeeze → squoze) (Cortese et al. 2006), suggesting that the memory traces for the stored regular forms are weakened as well, allowing responses reflecting competing irregular phonological patterns to emerge. All patients in these studies had cortical atrophy limited to the temporal lobe, particularly the left temporal pole, and were impaired at semantic memory tasks, including picture naming, fact retrieval, and synonym generation (Patterson et al. 2001; Tyler et al. 2004; Cortese et al. 2006; Patterson et al. 2006). Indeed, poor performance in synonym judgement predicted poor performance in irregular past tense production, but not in regular past tense production, consistent with a semantic basis for the irregular production impairment (Patterson et al. 2001; Patterson et al. 2006).

16.3.1.3 Herpes Simplex Encephalitis Herpes Simplex Encephalitis (HSE) is caused by a herpes simplex virus type 1 infection in the brain (Whitley 1990). The herpes virus follows a proscribed pathway through neural tissue, selectively damaging limbic system regions, including the insula and 6 The deficits were large enough that the concerns raised for Cortese et al.’s (2006) Alzheimer’s results (footnotes 3 and 4) are unlikely to have influenced the results for the Semantic dementia patients.

388

matthew walenski

orbitofrontal cortex, but particularly the temporal lobe, where the damage is greatest (Damasio and Van Hoesen 1985; Whitley 1990; Kapur et al. 1994). Motor cortex appears to be unaffected (Damasio and Van Hoesen 1985) and any damage to the basalganglia may be unilateral and relatively mild (Kapur et al. 1994). Cognitively, patients with HSE often have severe amnesia, behavioural disturbances, and loss of semantic memory, including naming deficits (Kapur et al. 1994; Barbarotto et al. 1996). English (Table 16.1). The elicited production of inflected forms has been examined in four patients with HSE (Tyler et al. 2002). All patients had impaired semantic memory and temporal lobe damage, particularly in inferior and lateral temporal cortex, the temporal pole, and medial temporal cortex in the left hemisphere, with sparing of left inferior frontal gyrus. Syntactic abilities were largely spared in all patients and no phonological impairments were noted (Warrington and Shallice 1984; Tyler et al. 2002). On real past tense forms, the patients were more impaired at the production of irregulars than consistent regulars. The patients were also impaired at the production of past tenses for novel consistent regular verbs. These impairments were not investigated in relation to the patients’ cognitive and linguistic profiles, and no information was provided on the errors they made. Thus while the pattern for real verbs is expected based on their impaired semantic memory, the source of the unexpected impairment at novel forms is not clear.

16.3.1.4 Summary The disorders discussed in this section reveal a characteristic performance pattern at the production of inflected morphological forms, such that disordered semantic memory and cortical temporal lobe damage leads to: (1) impaired production of real irregular inflected forms and a paucity of irregular patterns on novel verbs, with a high rate of memory-based errors (over-regularization errors in English; lexical or over-regularization errors in Italian); (2) impaired production of semi-regular forms (default class II and class III past participles in Italian); (3) impaired production of regular forms which are attracted to an irregular pattern, either phonologically (inconsistent regulars in English) or morphologically (novel class II present tense forms in Italian), with a high rate of memory-based errors for such forms in both languages (lexical errors in Italian; over-irregularized responses in English, particularly for Semantic dementia patients). The results are consistent with dual-system and singlemechanism claims of a dependence of the impaired forms on declarative (semantic) memory.

16.3.2 Disorders Affecting Frontal/Basal-ganglia Circuits and Motor Function The three disorders discussed in this section all involve dysfunction of frontal/basalganglia circuits and motor function. According to dual-system models, patients with

disorders

389

these disorders should display abnormal production of default affixed forms, with the abnormalities consistent with the nature of the motor dysfunction in the disorder (e.g. an inhibitory motor profile should correspond to inhibition of affixation). In addition, regular impairments are predicted to pattern with impairments to grammar (i.e. syntax and rule-based aspects of phonology). Single-mechanism accounts also predict impairments of regularized forms, but claim that any such regular impairment reflects an underlying phonological dysfunction.

16.3.2.1 Parkinson’s disease Parkinson’s disease is an adult-onset neurodegenerative disorder characterized particularly by impaired voluntary movement, manifesting as akinesia (impaired ability to initiate movement), hypokinesia (slow, reduced amplitude movements), and bradykinesia (slowed movements) (DeLong 1990; van Hilten et al. 1998; Berardelli et al. 2001). Parkinson’s disease begins with degeneration of dopaminergic neurons in the substantia nigra pars compacta of the basal-ganglia. The reduced output from the substantia nigra leads to increased inhibition of frontal cortical areas and the hypokinetic motor symptoms in Parkinson’s (DeLong 1990; Saint-Cyr 1995; DeLong and Wichmann 2009). Dementia develops in about 30 per cent of cases as the disease advances (Janvin et al. 2003; Aarsland et al. 2005; Janvin et al. 2006), although some non-demented Parkinson’s patients also show cognitive impairments (Janvin et al. 2003). With respect to language, syntactic deficits and lexical retrieval deficits have been observed in nondemented individuals (Murray 2000, 2008; Zanini et al. 2003). Phonological problems, including paraphasic errors and lack of prosody, have also been observed (Darkins et al. 1988; Zanini et al. 2003). Elicited production of inflected verb forms has been examined in English and Greek. English (Table 16.1). On average, patients performed similarly at consistent regulars, novels, and irregulars, and were impaired at all three verb types, although some patients showed no deficits at all (Ullman et al. 1997; Longworth et al. 2005; Ullman, in press). The apparently similar deficits in the different types of inflection, in conjunction with both motor and retrieval deficits in the disorder, underscores the importance of investigating performance in relation to other cognitive measures (e.g. motor function and memory), to determine whether there may be common or divergent sources for the similar impairments. Indeed, increases in right side hypokinesia (of hand and foot movements) corresponded to a greater deficiency at the production of real and novel regular past tenses, but not real or novel irregulars (Ullman et al. 1997; Longworth et al. 2005; Ullman, in press).7 In contrast, picture naming difficulties in 7 In Longworth et al.’s (2005) results, these effects were large but only approaching significance for the Ullman task (Cohen’s d = 0.77 for consistent regulars and 1.08 for novels). For the CSL task, the effect for consistent regulars was small (Cohen’s d = 0.10) and clearly not statistically reliable. Effect sizes were calculated based on the provided means and standard deviations of the hypokinesia scores for patients with and without impairment.

390

matthew walenski

the patients correlated with difficulties producing memory-dependent past tenses, but not rule-produced past tenses (Ullman, in press), consistent with a retrieval deficit affecting memory-dependent forms. The errors made by Parkinson’s patients also suggest a difficulty with affixation (Ullman et al. 1997; Longworth et al. 2005; Ullman, in press). The most frequent errors across all verb types were unmarked errors (look). Virtually no over-regularization errors (digged) were made to irregular verbs, in contrast to the pattern found for temporal-lobe damage, where over-regularization errors were frequent. A potential link to syntactic abilities was not investigated. In relation to phonology, hypokinetic Parkinson’s patients performed worse at producing consistent regular past tenses than phonologically matched irregulars (passed vs. lost), suggesting that phonological differences between regular and irregular forms did not contribute to the consistent regular deficit (Ullman et al. 1997; Ullman, in press). Examination of sex differences revealed that the pattern of hypokinesia correlating with performance at regulars and novels, and of picture naming performance correlating with performance at irregulars, held for men with Parkinson’s, whereas neither regular nor irregular past tense accuracy correlated with hypokinesia for women with Parkinson’s. Instead, accuracy for both types of verb correlated with picture naming scores in women with Parkinson’s (Ullman et al. 2002; Ullman et al. 2008). This pattern is consistent with other sources of evidence that women but not men memorize regular past tense forms—representing them in memory like irregulars (Ullman et al. 2002; Ullman et al. 2008; Prado and Ullman 2009). Greek (Table 16.3). Greek-speaking Parkinson’s patients produced real regular (sigmatic)8 and real irregular (non-sigmatic) past tense forms from a given present tense form (Terzi et al. 2005). On average, patients performed similarly on regular and irregular forms, and were impaired at both, although seven of the twenty-seven patients made no errors. Note that as regular and irregular forms have the same person– number agreement affixes in Greek, these affixes may be supplied by grammatical rule equally for regular and for irregular forms, potentially explaining an impairment at both in Parkinson’s (Terzi et al. 2005). However, without analyses relating motor (e.g. right-side hypokinesia) or memory function (e.g. picture naming) to performance at the production of regular and irregular forms, and without information about the errors made by the patients, the pattern in Greek is difficult to interpret. In relation to other areas of language, the same patients also exhibited a syntactic deficit (worse performance than controls at the comprehension of passive sentences); phonological performance was not examined (Terzi et al. 2005).

16.3.2.2 Huntington’s disease Huntington’s disease (HD) is a progressive condition characterized by motor, cognitive, and emotional disturbances (Kirkwood et al. 2001; Walker 2007). Neurologically, 8

One of their ten regular verbs was a semi-regular sigmatic form (poulo ‘sell’ → pouli-sa).

disorders

391

Table 16.3 Regular and irregular Greek perfect verb formsa Present, 1 sgb

Perfect, 1 sgb

Characteristics

First conjugation sigmatic (regular) aku-o ‘hear’ aku-s-a • perfect stem + -s suffix (+agreement) γraf-o ‘write’ e-γrap-s-a • default pattern applied to novel verbs • Predictable changes to stem-final C with addition of -s suffix Second conjugation sigmatic (semi-regular) mil-o ‘speak’ mili-s-a • Vowel-extended stem + -s suffix • Appended vowel is not predictable Non-sigmatic (irregular) vlep-o ‘see’ iδ-a • Perfect stem unpredictably distinct from present stem γern-o ‘bend’ e-γir-a • No -s suffix a Greek verbs are divided into three main categories—sigmatic first conjugation verbs, sigmatic

second conjugation verbs, and irregulars (Holton et al. 2007). Sigmatic verbs vastly outnumber non-sigmatic ones (roughly 2,119 to 147), and represent the pattern most frequently generalized to novel verbs (Stavrakaki and Clahsen 2009b). Monosyllabic stems (whether sigmatic or nonsigmatic) that begin with a consonant and have monosyllabic affixes in the perfect are obligatorily augmented by an initial e- to carry the stress (Holton et al. 2007; Ralli 2007). First conjugation verbs are stressed on the final syllable of their stem (penultimate syllable of the first person singular present tense) (Holton et al. 2007). They form the perfect stem by adding sigma (/-s-/; hence ‘sigmatic’) to the imperfect stem before the person/number agreement affix is added (/-a/). In verbs with imperfective stems ending in a consonant, the addition of the –s affix triggers application of a default phonological rule (specific to particular consonants), which mutates or deletes the stem-final consonant. For example, stem-final labial and velar fricatives change to plosives before -s (Ralli 2007). Second conjugation verbs are stressed on the final syllable of their first person singular present tense form, and also add sigma to the imperfect stem, but insert a vowel (-a-, -i-, or -u-) between the imperfect stem and the sigma. These verbs have been treated as semi-regular forms (i.e. an irregular stem with the regular perfective affixes) on some approaches (Tsapkini et al. 2002; Ralli 2007). However, they were included together with regular (first conjugation) verbs in two of the studies discussed here (Terzi et al. 2005; Stavrakaki and Clahsen 2009a). Non-sigmatic (irregular) forms do not add -s- to the stem. Instead, the imperfective stem changes unpredictably, sometimes with full suppletion. b Verbs are shown with the first person singular agreement affix. The same person/number agreement affixes are used for regular, semi-regular, and irregular forms for present and perfect active verb forms.

Huntington’s is associated with the loss of neurons in the neostriatum, particularly the caudate nucleus and putamen. This degeneration gives rise to the motor symptoms in the disorder, which are primarily characterized by unsuppressible, hyperkinetic ‘choreic’ movements, although hypokinetic movements and rigidity are also found, even in the same individuals who exhibit choreic movements (Reiner et al. 1988; DeLong 1990; Young and Penney 1993). Cognitively, Huntington’s patients have impairments of procedural, episodic, and semantic memory (Hodges et al. 1990; Saint-Cyr et al. 1995; Solomon et al. 2007).

392

matthew walenski

In contrast to the temporal-lobe disorders however, the episodic and semantic memory deficits in HD appear to reflect difficulty in retrieving stored information, rather than the degradation or loss of this information. With respect to language, impairments of syntax and lexical retrieval (e.g. naming) are frequently found, even in early stages of the disease (Podoll et al. 1988; Hodges et al. 1990; Murray 2000). With respect to phonology, there are indications of a deficit perceiving prosodic contours and phonemes in sentences (but not in single words) (Speedie et al. 1990; Teichmann et al. 2009). English (Table 16.1). On average, patients with Huntington’s disease inflected irregular (dug), regular (looked), and novel (plagged) verbs at similar rates of success, and with lower accuracy than controls for all three verb types (Ullman et al. 1997; Longworth et al. 2005; Ullman, in press). Two types of errors predominated across all three verb types—unmarked forms and affix errors, including multiply-suffixed (diggeded, lookeded, plaggeded) responses and over-regularization errors to irregular verbs (digged).9 The errors revealed an interesting relationship to motor function in the disorder (Ullman et al. 1997; Ullman, in press). The rate of unmarked errors correlated with right-side hypokinesia for regulars and novels, but picture naming for irregulars. In contrast, the rate of affix errors correlated with chorea for regulars, novels, and irregular verbs. This suggests that the impaired group average performance seen for all three verb types may have reflected the joint influence of three factors: hyperkinesia, hypokinesia, and a retrieval deficit. French (Table 16.4). Three groups of French speaking Huntington’s patients were given the infinitive forms of real regular verbs, real irregular verbs, novel regular first conjugation verbs (e.g. garouster; based on the pattern arriver ‘arrive’—elle arrive ‘she arrives’—elle arrivera ‘she will arrive’), and novel ‘sub-regular’ verbs (e.g. saurentir, olissoire; based on the patterns finir ‘finish’—elle finit ‘she finishes’—elle finira ‘she will finish’ and croire ‘believe’—elle croit ‘she believes’—elle croira ‘she will believe’),10 and tested at the elicited production of third person singular present tense and future tense forms (Teichmann et al. 2005; Teichmann et al. 2008; Teichmann et al. 2009). Relative to unimpaired controls, patients were consistently spared across studies at the production of real regular forms, but consistently impaired at both sets of novel verbs and real irregulars, with the most severe impairments at the novel ‘sub-regular’ verbs. Errors were due to excess suffixation (over-regularizations saurent-era; double suffixed forms saurent-ir-era), incorrect suffixes (saurent-ure) and repetitions of the given form (saurentir). The novel ‘sub-regular’ verbs had the highest rate of excessively suffixed

9 Longworth et al. (2005) found a lower rate of regularization errors to irregular verbs than Ullman et al. (1997) and no multiply-affixed errors. 10 While croire is an irregular verb, and its third person singular present and future tense forms are generated from a lexically specified stem, the infinitive is generated from this same lexically specified stem. Thus the transformations from the infinitive (given to the participant) to these inflected forms could be accomplished without reference to lexical knowledge, even for real verbs. Note that finir is a completely regular verb.

disorders

393

Table 16.4 French present and future tensea Infinitive

Present Future Characteristics tense (3ps) tense (3ps)

laver ‘wash’

lave

lavera

• All forms produced from a single lexical stem • All forms are regular

Second conjugation Regular finir ‘finish’

finit

finira

• Same as first conjugation verbs

met boit vient va

mettra boira viendra ira

• Irregular verbs have more than one lexically specified stem • Forms built from these additional lexically specified stems are irregular • Default phonological relations hold between stems that are not lexically specified

First conjugation Regular

Third conjugation Irregular

mettre ‘put’ boire ‘drink’ venir ‘come’ aller ‘go’

a French verbs are traditionally divided into three conjugations. First conjugation verbs have infini-

tives in -er, and are completely regular. The majority of verbs in French belong to the first conjugation, and it is the most productive class (Meunier and Marslen-Wilson 2004). Second conjugation verbs have infinitives in -ir, and have -iss- throughout the imperfect. The second conjugation has fewer members than the first, and is only marginally productive (Harris 1987). Third conjugation verbs have infinitives in -re and -oir, as well as verbs with -ir infinitives that do not have -iss- in the imperfect, and the irregular verbs with -er infinitives (e.g. aller ‘to go’). This class is not productive, and has many irregular members. Note that in this pedagogical classification all irregular verbs are assigned to the third conjugation. On a recent linguistic treatment, verb classes are entirely done away with and regularity is defined instead by the number of stems that need to be lexically represented in order to allow for correctly generating all of a verb’s inflected forms (Boyé 2000; Bonami and Boyé 2002). On this view, traditionally first and second conjugation verbs like laver and finir are entirely regular— each verb has only a single lexical stem (e.g. /lav-/). Additional stems are either identical to this default stem (specified as the stem used for imperfect forms and the 1st and 2nd plural present tense forms), or are related to it by default phonological transformations. Irregular verbs are those which have additional lexically specified stems, which cannot be related to the verb’s other stems by productive phonological rules (e.g. boire has three distinct stems: default /byv/ and additionally specified /bwa/ and /bwav/). Note that Bonami and Boyé (2002) use the term ‘semiregular’ to refer to a verb with two lexical stems, and ‘irregular’ to refer to verbs with three lexical stems (note that this differs from the use of semi-regular and irregular in this chapter, which refers to the form itself, not the entire verb). For all verbs, inflected forms are generated by addition of the default affixes and phonological transformations (except for a small number of fully suppletive forms, such as for etre ‘be’). Irregular forms can be defined as those which are generated from the non-default, lexically specified stems. On this view, the third person singular present tense is formed by adding a null suffix to the stem; the third person singular future tense is formed by adding the suffix -ra to the stem.

394

matthew walenski

responses. Syntactic deficits were also consistently found for comprehension of noncanonically ordered sentences (passives and object-relative clauses, in which the patient argument precedes its verb), but less consistently (and less severely) for canonically ordered sentences (actives and subject-relative clauses, where the arguments follow canonical agent–verb–patient order). Performance in elicited production and syntax comprehension was also examined in positron emission tomography (PET) scans of resting brain metabolism (Teichmann et al. 2008). Performance in novel verbs (both types) correlated with left ventral striatum and ventral portions of the left caudate head. Performance in novel regular first conjugation verbs was additionally linked to ventral regions of the left putamen. Similarly, comprehension of non-canonical sentences was linked to bilateral ventral striatum, and ventral regions of the left putamen. In contrast, performance in irregular verb inflection and comprehension of canonical sentences showed a different pattern: both correlated with dorsal portions of the left caudate head (Teichmann et al. 2008).

16.3.2.3 Tourette’s syndrome Tourette’s syndrome is a developmental disorder defined by verbal and motor tics (APA, 1994). The expression of tics is fast and involuntary, and the tics appear to be caused by abnormalities of the basal-ganglia and closely connected regions of cortex, especially motor and cognitive regions of the frontal lobes, and abnormal levels of dopamine (Bradshaw 2001; Albin 2006; Kienast and Heinz 2006; Mink 2006). These abnormalities are thought to result in decreased inhibition of frontal activity, leading to a hyperkinetic behavioural profile and an inability to suppress tics (Osmon and Smerz 2005; Albin and Mink 2006). Cognitively, Tourette’s patients are impaired on at least some procedural memory tasks (Stebbins et al. 1995; Keri et al. 2002; Channon et al. 2003; Marsh et al. 2004; Marsh et al. 2005), but appear to be spared on tasks of declarative (semantic) memory (Stebbins et al. 1995; Keri et al. 2002; Channon et al. 2003; Marsh et al. 2004). Language abilities in Tourette’s remains essentially unexplored, aside from studies of vocal tics (Legg et al. 2005) and a single study of inflectional morphology (Walenski et al. 2007). English (Table 16.1). Children with Tourette’s were abnormally fast at producing rule-governed past-tenses: consistent regular verbs (slipped), regularized responses to novel regular past tenses (plimmed), and over-regularization errors to real irregulars (bringed) (Walenski et al. 2007). In contrast, the children with Tourette’s were not quick at producing memory-dependent past tenses: real irregulars (brought), inconsistent regulars (squeezed), and irregularized responses to novel irregular verbs (splam). The children with Tourette’s showed a small drop in accuracy at the production of real consistent regular forms, but otherwise the proportions of correct responses and specific types of errors (whether quicker than normal or not) were no different than for matched control participants. Multiple measures of potential phonological complexity

disorders

395

differences between the regular, irregular, and novel verbs were controlled in the study, and did not explain the response time or accuracy pattern.11

16.3.2.4 Summary The three disorders involving frontal/basal-ganglia dysfunction discussed here all show abnormal production of regular affixed responses. The nature of the abnormalities reflects the motor profile of each disorder: hypokinesia in Parkinson’s and Huntington’s is correlated with the omission of default affixes; hyperkinesia in Huntington’s is correlated with excess affixation; and children with Tourette’s (who by definition have tics) show quick responses for -ed affixed responses. Production of memorydependent forms appears to be related to a retrieval deficit in Parkinson’s disease. In Huntington’s disease, observed irregular deficits seemed to reflect both chorea (related to erroneous addition of an affix) and a retrieval deficit (related to unmarked errors). In relation to other aspects of language, evidence from Huntington’s suggests a link between regular affixed forms and complex syntax, consistent with dual-system claims. In contrast, phonological factors do not appear to be sufficient to explain the abnormal production of regular morphological forms in Parkinson’s and Tourette’s, contra single-mechanism claims.

16.3.3 Additional Disorders This group of disorders each involve multiple domains of neurocognitive dysfunction. Nevertheless, the predictions made by dual-system and single-mechanism models are the same as for other disorders: the declarative/procedural model predicts that impairments at producing memory-dependent inflected forms should correspond to temporal-lobe damage and semantic memory dysfunction, whereas impairments at producing affixed regular and novel forms should correspond to frontal/basalganglia damage as well as motor dysfunction and syntactic impairments. The singlemechanism account predicts correspondences between irregulars, temporal-lobe damage, and semantic memory dysfunction on the one hand, and frontal lobe damage, phonological dysfunction, and regulars and novels on the other. 11

Six of thirty-one of the consistent regular past tenses were phonologically complex per Burzio’s (2002) criteria (a long vowel followed by a consonant cluster; drowned, failed, rolled, sailed, scraped, and signed), compared to only two of thirty of the irregulars (sold, told), but eight of sixteen inconsistent regulars (cleaned, climbed, heaped, pleased, screamed, screened, squeezed, wiped). Thus while the stimuli are not perfectly matched, it is implausible to suggest that this factor both underlies a drop in accuracy for consistent regular past tenses and at the same time also underlies quick response times for these verbs. Even if this were somehow plausible, it would remain problematic that the inconsistent regulars patterned with the irregulars for both accuracy and response time, despite the more severe mismatch in phonological regularity relative to the consistent regulars and irregulars.

396

matthew walenski

16.3.3.1 Williams syndrome Williams syndrome is a rare genetic disorder caused by the deletion and mutation of genes on chromosome 7q11.23. Cognitively, individuals with Williams syndrome have good auditory rote memory, poor visuo-spatial abilities, an overfriendly, highly social demeanour, and generally very low IQ (Bellugi et al. 1994; Brock 2007). With respect to language, phonological skills appear to be almost completely typical (Brock 2007). The production and comprehension of sentences, even those with complex syntax, are also relatively strong areas, although abilities in this area may be worse than those of chronological age-matched peers (Brock 2007). Performance on receptive vocabulary tasks is also a relative strength, whereas performance on production tasks (e.g. picture naming) appears to be abnormal (Clahsen and Almazan 1998; Brock 2007). Elicited production of inflected verb forms has been examined in English, Greek, and German. English (Table 16.1). Individuals (children and adults) with Williams Syndrome were impaired at the production of real and novel English irregular past tenses, when compared with control groups carefully matched for mental age, but were relatively spared at producing consistent regular, denominal, and novel regular past tenses (Bromberg et al. 1994; Clahsen and Almazan 1998; Thomas et al. 2001; Clahsen and Temple 2003; Clahsen et al. 2004). In addition, as for matched typically-developing controls (but unlike children with specific language impairment), production of consistent regular past tenses in Williams syndrome was not affected by the phonotactics of the verb stems (Marshall and van der Lely 2006). Over-regularization errors to irregular verbs were consistently produced at high rates, with some unmarked errors as well (Bromberg et al. 1994; Clahsen and Almazan 1998; Thomas et al. 2001; Clahsen and Temple 2003; Clahsen et al. 2004). In comparison to other areas of language, the same children who were spared at producing regular forms had adult-like comprehension of passive sentences and pronouns (Clahsen and Almazan 1998). Greek (Table 16.3). In Greek, children with Williams syndrome performed more accurately than matched controls at producing sigmatic (regular) past tenses from third person singular present tense forms (note that the regular stimuli included regular and semi-regular verbs), but less accurately at producing non-sigmatic (irregular) past tenses (the stimuli included stem-changing and suppletive irregulars) (Stavrakaki and Clahsen 2009a). In a separate study, children with Williams syndrome were again more accurate than matched controls at producing regular sigmatic past tenses (γraf-o → eγrap-s-a), semi-regular sigmatic past tenses (mil-o → mili-s-a), and irregular (stem-changing non-sigmatic) past tenses (γern-o → eγir-a), but slightly less accurate than controls at producing suppletive non-sigmatic past tenses (vlep-o → iδ-a) (Varlokosta et al. 2008). For novel verbs, children produced sigmatic past tenses substantially more often than controls in one study (Varlokosta et al. 2008), but not the other (Stavrakaki and Clahsen 2009a). German (Table 16.5). German-speaking teenagers with Williams syndrome had essentially normal productive syntax and were completely unimpaired when producing

disorders

397

Table 16.5 Regular and irregular past participles in Germana Infinitive

Past participle

Characteristics

spielen ‘play’ bemerken ‘notice’

ge-spiel-t bemerk-t

• stem + default -t suffix • stem is the same as the stem of the infinitive • default pattern

denken ‘think’

ge-dach-t

• stem is changed from stem of the infinitive • adds the default –t suffix

rufen ‘call’ verzeihen ‘forgive’ schneiden ‘cut’ beschliessen ‘decide’

ge-ruf-en verzieh-en ge-schnitt-en beschloss-en

• stem may or may not be the same as that of the infinitive • stem changes are unpredictable • stem adds a non-default –n suffix

Regular

Semi-regular

Irregular

a Regular past participles in German are formed by adding the suffix –t to the verb stem. Irreg-

ular past participles add the suffix –n, and may involve changes to the stem as well. It is not predictable from the infinitive whether a particular verb has a regular or irregular past participle (Penke and Krause 2004). Semi-regular past participles (also referred to as ‘mixed’ forms) add the -t suffix to an irregular stem. Only thirteen verbs have semi-regular past participles formed this way. The –t suffix is the default, and is used for new verbs entering the language, although it is not the most frequent pattern (either by type or token). The ge- prefix is added to both regular and irregular past participles when the verb stem is stressed on the first syllable, and is therefore phonologically conditioned, except that verbs borrowed from foreign sources that end in –ieren or –eien do not add the ge- prefix regardless of their stress pattern (all such verbs are regular). Inseparably prefixed verbs, which are never stressed on the first syllable, do not add this prefix. For further details, see Clahsen (1999) and references therein.

regular past participles from first person singular present tense forms, in comparison to chronological-age matched and mental-age matched controls (Penke and Krause 2004). The teenagers with Williams syndrome performed less accurately than chronological-age matched controls when producing irregular past participles (but similarly to mental-age matched controls). However, the children with Williams syndrome made more over-regularization errors on frequent irregular participles than even mental-age matched controls. Furthermore, unlike for the controls, whose proportion of such errors went down with increasing mental age (as expected, as increased exposure to irregular forms over time should improve accuracy), the participants with Williams syndrome showed a constant rate of such errors, regardless of mental age (Penke and Krause 2004). For novel past participles, the Williams participants were no different than controls in the proportion of responses with the regular -t, but produced slightly fewer irregularized -n responses (one, vs. 6 for the controls) (Penke and Krause 2004). However, in an independent case study, a 12-year-old girl with Williams

398

matthew walenski

syndrome had near perfect performance at passive comprehension as well as at both regular and irregular past participle production, although no information about the stimuli or tasks was given (Schaner-Wolles 2004).

16.3.3.2 Autism spectrum disorders Autism is a developmental disorder that is characterized by abnormal social interaction, abnormal language and communication, and restricted and repetitive behaviours and interests (APA, 1994). The biological basis of autism is unknown, although growing evidence suggests a genetic basis for the disorder (Geschwind 2011). Anatomically, multiple regions, if not the entire brain, appear to develop abnormally in autism. Neuroimaging studies reveal functional and structural abnormalities of frontal/basalganglia and temporal-lobe regions that relate to language deficits, although in studies that did not measure language or cognitive abilities, structural abnormalities of these regions have not been found consistently (for review, see Walenski et al. 2006; Amaral et al. 2008). Cognitively, evidence suggests that procedural memory is impaired or abnormal in autism (Mostofsky et al. 2000; Ullman 2004; Walenski et al. 2006). Likewise, episodic memory appears to be consistently impaired in the disorder, whereas aspects of semantic memory appear to be largely spared, at least in high-functioning autism (Bennetto et al. 1996; Minshew and Goldstein 2001; Ben Shalom 2003; Crane and Goddard 2008). Note that it has been argued that a different profile holds for low-functioning autism, with impaired episodic and semantic memory, but spared procedural memory (Boucher et al. 2008). In language, pragmatic deficits are ubiquitous in autism, as are prosodic abnormalities (for review, see Shriberg et al. 2001; Walenski et al. 2006; Diehl et al. 2008). Impairments of syntax are also found, although not necessarily in all individuals (Walenski et al. 2006; Eigsti et al. 2007; Williams et al. 2008; Eigsti and Bennetto 2009; Perovic et al. 2012). In contrast, performance at tasks involving single words, in both receptive and expressive domains, tends to be relatively spared, and may even be enhanced in some respects, particularly for high-functioning individuals (Walenski et al. 2006, 2008; Luyster and Lord 2009; Norbury et al. 2010). Elicited morphology has been investigated in a small number of studies in English. English (Table 16.1). A computer-based study of elicited past tense production found no differences in accuracy scores or error types for consistent regular, irregular, inconsistent regular, and novel verbs between high-functioning boys with autism and typically developing control boys, controlling for age and IQ (among other factors) (Walenski et al. 2006, 2014). In an earlier study, children with autism were tested at the elicited production of present tense and past tense verb forms. Language-impaired children with autism were compared with age-matched groups of children with autism who either had unimpaired language or were borderline impaired, where language impairment was defined by performance on a receptive vocabulary test (Roberts et al.

disorders

399

2004). Relative to the unimpaired and borderline groups, whose past tense performance did not differ, the language-impaired group produced fewer correct regular forms and fewer correct irregular forms, with high rates (>20 per cent) of unmarked errors to both. Across all subjects, peformance in regular and in irregular past tense production was correlated with non-verbal IQ; however, performance was not investigated with respect to other measures (e.g. of lexical retrieval) for regular and irregular verbs separately, and so the source of the deficits on the two verbs types is not certain. In addition, comparisons with age- and IQ-matched typically developing controls were not made, so it is not clear if the borderline or unimpaired groups may have also shown impairments, either at regular verbs, irregular verbs, or both. Performance on third person singular present tense forms was also impaired for the language-impaired group, but results were not reported separately for regular and irregular forms, rendering their interpretation uncertain. Two additional studies that report deficits at past tense production also did not report regular and irregular performance separately, and so are not discussed (Bartolucci and Albers 1974; Botting and Conti-Ramsden 2003). Examination of response times (Walenski et al. 2006, 2014) revealed that the high-functioning boys with autism were abnormally fast at producing affixed forms, with quick response times for consistent regular past tenses (slipped), regularized responses to novel regular verbs (plimmed), and over-regularization errors to real irregulars (bringed), but not for irregular past tenses (brought), inconsistent regulars (squeezed), or irregularized responses to novel verbs (splam). Links between morphology and syntax were not examined, although phonological differences between the regular, irregular, and novel verbs were controlled (including controlling for phonological irregularity in the data analysis; the same stimuli were used as for the study of Tourette’s syndrome discussed in Section 16.3.2.3), and so potentially abnormal phonological processing in autism (whether enhanced or impaired) is unlikely to explain the observed pattern. The similarity of the quick response time pattern to that found in Tourette’s (see Section 16.3.2.3) suggests the involvement of basal-ganglia abnormalities in autism for -ed suffixed forms.

16.3.3.3 Schizophrenia Schizophrenia is a neurological disease characterized by positive symptoms (delusions, hallucinations), negative symptoms (loss of motivation, apathy, anhedonia, avolition, poverty of speech, etc.), disorganization of thinking and behaviour (‘thought disorder’), and disruption of mood and cognitive function (Tandon et al. 2009). Symptom severity in these areas is heterogeneous over time and across individuals (Keshavan et al. 2011). With respect to memory, patients with schizophrenia perform worse (i.e. with slower, less accurate responses) than healthy controls on tasks tapping the procedural memory system, even when general cognitive abilities are taken into account, although

400

matthew walenski

procedural learning itself appears to be spared (Keri et al. 2000; Weickert et al. 2002). Conversely, declarative knowledge learned early in childhood (prior to disease onset) appears to remain essentially normal, whereas deficits are found in learning new declarative knowledge (Dalby and Williams 1986; Goldberg and Weinberger 1996; Weickert et al. 2000; Nuechterlein et al. 2004). Neurologically, dysregulation of dopamine in prefrontal cortex and connected subcortical structures, including the basal-ganglia, is one of the clearest and most reliable findings (Goldman-Rakic, et al. 2004; Keshavan et al. 2011). In language, many studies report deficits at both syntactic and lexical processing, in both production and comprehension (for reviews, see DeLisi 2001; Covington et al. 2005). However, while lexical knowledge may become disorganized and/or difficult to access, the knowledge itself does not appear to be affected (Covington et al. 2005). Phonological deficits have also been reported (Crow et al. 1995; Angrilli et al. 2009). Morphology has received brief mention in only a handful of studies (Kleist 1914; DeLisi et al. 1997; Covington et al. 2005), and inflection specifically has only been investigated in one (Walenski et al. 2010). English (Table 16.1). Patients with schizophrenia were spared at producing irregular English past tenses but impaired in the production of regular and novel past tenses (Walenski et al. 2010). Incorrect responses to regular verbs were predominantly unmarked forms (shrug) and lexical intrusions (e.g. shrug → guessed); incorrect responses to novel verbs were predominantly lexical intrusions (plag → fished). While syntactic abilities of the patients were not examined, phonological differences between the regular, irregular, and novel verbs were controlled, and so a phonological deficit in schizophrenia is unlikely to explain the observed pattern (Walenski et al. 2010).

16.3.3.4 Summary The three disorders discussed in this section are complex, each having deficits in multiple cognitive domains. Nevertheless, their production of inflected forms is consistent with the patterns displayed by the better-understood and more focal disorders. Both schizophrenia and (high-functioning) autism conform to the frontal/basal-ganglia disorder profile. Performance in both disorders is consistent with dual-system predictions, and problematic for single-mechanism claims, as phonological factors did not explain the findings. In Williams syndrome, the results across three languages are similar to the characteristic pattern exhibited by patients with a semantic memory disorder. Dual-system models therefore expect an underlying semantic deficit to account for the irregular deficit in Williams syndrome. Proponents of single mechanism accounts have argued that the performance in Williams syndrome instead reflects an overly strong focus on phonology, such that the resulting imbalance between phonology and semantics yields weakened semantic representations and poor performance at the production of irregular inflected forms (Thomas et al. 2001) (for arguments against this position, see Clahsen and Temple 2003).

disorders

401

16.4 Conclusions

.............................................................................................................................................................................

The evidence discussed here, while limited in scope to elicited production and verbs, from a relatively small subset of possible disorders and languages, nonetheless elucidates the broader theoretical perspectives concerning the biological and neurocognitive basis of inflection. From this perspective it can be observed that patients with similar profiles of neurological damage and cognitive deficits also show similar patterns of deficient performance at the production of inflected verb forms. In one profile, damage to temporal-lobe cortical regions that results in a declarative (semantic) memory deficit also leads to an impairment of memory-dependent inflected forms, including irregulars, certain regulars, and semi-regulars. The disorders with this profile include Alzheimer’s disease, Semantic dementia, Herpes Simplex Encephalitis, and Williams syndrome. In the second profile, exemplified here by Parkinson’s disease, Huntington’s disease, Tourette’s syndrome, high-functioning autism, and schizophrenia, damage to frontal/basal-ganglia circuits leads to motor and cognitive dysfunction, including commonly in syntax and phonology, and abnormal production of default-affixed regular forms. The dysfunction of regular morphology in these disorders appears to be linked to syntactic but not phonological deficits. Similar patterns of deficient inflection should also be found in other patient populations with neurocognitive profiles similar to those discussed here. While discussion of additional disorders is beyond the scope of this chapter, two that have received a great deal of attention in the literature deserve brief mention. Specific Language Impairment (SLI) is a developmental disorder defined by language impairments in the absence of any clear mental or physical handicap or frank neurological damage (Bishop 1992; Ullman and Pierpont 2005). Studies across a variety of languages suggest that children with SLI conform to the frontal/basal-ganglia pattern, with motor deficits as well as deficits in syntax, phonology, and morphology (for review and discussion, see Gopnik 1997; Leonard 1998; Hansson and Leonard 2003; Joanisse and Seidenberg 2003; Ullman and Pierpont 2005; Marshall and van der Lely 2006). Aphasia is an acquired language impairment subsequent to brain damage, particularly to the left hemisphere. While there is a great deal of variability in the resulting language deficits, distinct aphasic syndromes have been identified (Goodglass 1993). Non-fluent aphasia (also referred to as anterior aphasia) has been argued to exhibit particular deficits in producing real and novel regular inflected forms, whereas fluent aphasia (also referred to as posterior aphasia) has been argued to show particular deficits in the production of irregular inflected forms (for review and discussion from this perspective, see Ullman et al. 1997; Bird et al. 2003; Ullman et al. 2005; Druks 2006; for a broader cross-linguistic discussion of morphology in aphasia, see Jarema 1998). Finally, while a great deal of progress has been made in identifying and explaining the neurocognitive basis of regular, irregular, and semi-regular inflected verb forms in elicited production, English is still the dominant language of investigation, and inflection remains almost completely unexplored (with any experimental paradigm)

402

matthew walenski

across numerous typologically distinct languages in these and other disorders. Thus, while the patterns of deficient performance discussed here appear to hold across patient populations and universally across languages, on this topic our current state of knowledge is certain to be far from the final word.

Acknowledgements

.............................................................................................................................................................................

I would like to thank Michael Ullman for his advice and support, and Harald Clahsen, Spyridoula Varlokosta, Anne-Catherine Bachoud-Lévi, Angela Ralli, and Martina Penke for graciously answering questions, and Mariel Pullman and Rachel Smith for their much appreciated assistance.

p a r t vi ........................................................................................................

SKETCHES OF INDIVIDUAL SYSTEMS ........................................................................................................

chapter 17 ........................................................................................................

VERBAL INFLECTION IN IHA: A MULTIPLICITY OF ALIGNMENTS ........................................................................................................

mark donohue

17.1 Iha

.............................................................................................................................................................................

Iha is a Trans New Guinea language from the western tip of New Guinea. It is spoken by at least 6,000 people living on the Onin peninsula in and around the city of Fakfak. Iha is the lingua franca of the Onin area, and has developed an isolating variant used by non-Iha people and when in communication with them; the variety described here is the mother-tongue of the inner north of the peninsula. In the following sections I shall describe the alignments coded by the different affixes, and the ways in which they combine. Iha has seven vowels (i e ε a ɔ o u), contrastive prosody, and the following consonants: p t̪ q kp d̪ mb n̪d̪ ng ŋmɡb m n̪ β ʁ ћ ɽ w j (Narfafan 2009 is a discussion of the segmental phonology). These are represented with their IPA symbols with the following exceptions, accommodating towards local orthographic preferences: dental articulation is not marked, the uvular consonants are transcribed as velar (and the velar nasal element as ), the voiced fricatives as and , the labialvelars as and , the liquid as and the palatal glide as . Syllable-final p, t, and k are written as , respectively, and the high vs. low mid vowels are not distinguished. Most CC sequences are broken up with an epenthetic vowel; in (3a) the verb whómbon is pronounced [wˇɔh´ɔmbɔn]. Similar epenthesis is found with word-initial j: /jaɽ/ ‘year’ (a loan from Dutch jaar) is pronounced [ı ’̯ ja:ɽ]. Tone is not generally marked, as in the orthography, since a discussion of morphotonemics is beyond the scope of the present chapter, although it is worth noting that, in addition to the suffixal segmental forms of

406

mark donohue

the tense–aspect morphemes, some tense–aspect suffixes contribute tone to the grammatical word as well; some lexical high pitches are indicated with an accent, but the high pitch associated with (for example) on in on-ya (example (3a)), which occurs as a result of the tone of -ya spreading over the inherently low-tone morpheme, is not shown. Similarly, not all morphophonemic alternations are explicated or discussed.

17.2 Inflectional Categories and Alignment ............................................................................................................................................................................. Agreement on the verb, coding participants in the clause, is usually restricted according to the role of the arguments or other participants that it codes.1 The categories that are relevant to the coding of arguments in verbal inflection are not necessarily those found in other sub-parts of the grammar, such as the pronominal case system, or the word order. The Nepali examples in (1) and (2) illustrate the fact that while agreement on the verb consistently indexes the subject, with -chu marking 1sg and -chau for 2sg, case marking on the subject NP follows an ergative pattern, with only As so marked, never Ss. There are thus two different alignment systems in the one language, one marked inflectionally on the verb and one marked by case on the arguments of the verb.2 Such a split, between ergative NPs and nominative verbs, has been widely attested across the world. (1) Nepali a. Ma 1sg ‘I stay.’

bas-chu. stay-1sg

(2) Nepali a. Maile ghiu dekh-chu. 1sg.erg ghee see-1sg ‘I see some ghee.’

b. Timi bas-chau. 2sg stay-2sg ‘You stay.’ b. Timi-le ghiu dekh-chau 2sg-erg ghee see-2sg ‘You see some ghee.’

In Iha we similarly find a split in the alignment system. Unlike better-documented cases (such as Nepali, or other languages in Indonesia reported in Donohue and Brown 1999), the split(s) in alignment marking in Iha all occur within the verb.3 Inflection on the verb refers to (a) a nominative category; (b) an undergoer category; (c) an 1 Siewierska (2005) reports eleven languages in a global sample of 380, 3%, that show ‘hierarchical’ marking, in which syntactic or semantic roles do not play an important role in the form of the agreement marking on the verb. 2 The following special abbreviations are used: 12: first person inclusive. ny, te: ny- and te-conjugations. 3 Iha possesses a case system that distinguishes ‘nominative topic’, beneficiary, instrumental ablative, locative, allative.

verbal inflection in iha: a multiplicity of alignments

407

accusative category, and (d) an intransitive argument category. Additionally, beneficiaries and some other arguments may be marked on the verb with a set of beneficiary suffixes. Outside the verb we also find a rich case system with nominative and semantic alignments, but will not explore this in detail here. In describing alignment, I make use of the terminology, following Comrie (1978), that describes different configurations of syntactic roles in theory-neutral terms, using three basic descriptors: A: the most agent-like argument of a bivalent verb; S: the single argument of a monovalent verb; P: the most patient-like argument of a bivalent verb. In addition to these groups, a more semantically-based description is sometimes necessary. Donohue (2008) and Wichmann (2008) describe the range of languages with ‘semantic alignment’, in which the different arguments of a verb are coded according to semantic, not syntactic, principles. It is entirely possible for semantic factors to influence the morphosyntactic properties of a syntactic role. Languages in which the S argument is differently coded according to whether it is more or less agentive or patientive, or in which the A is marked differently when more highly agentive, are common around the world.

17.3 Iha Verbal Agreement

.............................................................................................................................................................................

The simple verb in Iha4 can be described in outline with the following basic template; the positions we shall focus on, where pronominal inflection can be found, are shown in bold. In addition to the positions described here, additional aspectual, directional, and evidential positions can be added. A single verb illustrating all of these morpheme positions has not been attested, but (3) shows a pair of rather full examples of different inflectional categories on the verb. In (3)b the inflected verb is the whole string newepymyannyatnanweradembihtpyebon, including the evidential portium which is not described in the template in Figure 17.1. (3) Exuberantly-inflected verbs in Iha a. Ki-ndo-ko-qpag-qpen-ni-mb-ih.5 2-trans-recip-chase-1sg.ben-conj-ypst-3 ‘They made you hit each other for me.’ 4 5

Using ‘simple’ to refer to the fact that there is only one verb root, and no auxiliary TAM marking. In a phonologically satisfying solution, the labio-velar approximant /w/ is realized as a labial-velar stop [kp] following an obstruent; conversely, the complex labialvelar phoneme /ŋmgb/ is realised as [w] ˜ intervocalically in fast speech (eg. /ni̪ ŋmɡbit̪/ ‘afternoon’ is realised as [ni̪ wit̪ ˜ ]).

408

mark donohue

-4 AgrPat ni-

-3 Trans ndo-

-2 Recip ko-

-1 Accom kono-

0 √ root

+1 Benefactive -wen

+2 Conj -ni

+3 Tense -mb

+4 AgrNOM -e

fig. 17.1 Templatic structure of the Iha verb

b.

Mi-ge koret 3-foc food newe-py-myan-nyatna-n-wera-de-mb-ih eat-repet-direct-like.that-conj-very.many-cont-ypast-3 -t-pye-b-o-n pya. -evid-repet-12past-12-1sg assert ‘He kept on eating all sorts of things, for an extended time, again and again, and I witnessed it multiply and repeatedly, so do not doubt me.’

The following sections will examine these three inflectional categories, with an eye to examining the different alignment systems observed.

17.3.1 Final Suffixes All realis Iha verbs contain a portmanteau agreement/tense–aspect suffix. The features distinguished for agreement are the same in all tense–aspect categories (except the irrealis, which does not show agreement in this way), and will be illustrated in the main with yesterday’s past forms. The suffixes show agreement for and S or an A argument; that is, they index a nominative (i.e. S and A) grouping of arguments. Examples (4) and (5) demonstrate the contrasts available for this tense for monovalent and bivalent verbs, and show that, as in the Nepali example (1), there are no differences between the monovalent and bivalent verbs. Note that there is no number contrast for third persons, and that 1excl.pl and 1incl.pl are conflated in these suffixes. (4) Monovalent verb a.

b. c. d. e. f.

On-a Qpagqpag-na whó-mb-on. 1sg-top Fakfak-all go-ypst-1sg ‘I went to Fakfak yesterday.’ Ko-ya Qpagqpag-na whó-mb-e. Mi-ya Qpagqpag-na whé-mb-ih. In-a Qpagqpag-na whó-mb-ob. Mbi-ya Qpagqpag-na whó-mb-ob. Ki-ya Qpagqpag-na whó-mb-i.

2sg 3sg/3pl 1excl.pl 1incl.pl 2pl

verbal inflection in iha: a multiplicity of alignments

409

(5) Bivalent verb a.

b. c. d. e. f.

On-a mbyar knó-mb-on. 1sg-top dog see-ypst-1sg ‘I saw a/the dog yesterday.’ Ko-ya mbyar knó-mb- e. Mi-ge mbyar kné-mb-ih. In-a mbyar knó-mb-ob. Mbi-ya mbyar knó-mb-ob. Ki-ya mbyar knó-mb-i.

2sg 3sg/3pl 1excl.pl 1incl.pl 2pl

17.3.2 Verbal Agreement: Prefixes Verbs also show agreement for local persons (that is, the first or second persons implicit in a speech act) by prefix, as seen in (6)–(7). These examples make it clear that, in a bivalent verb, it is the P that is coded by prefix, and that number is not relevant to the form of the prefix. We can see from these examples that these two prefixes code different sets of features: ni- is used when the argument it indexes contains first persons, but no second persons, while ki- is used for all arguments with a second person participant, regardless of whether or not a first person is also included.6 This means that while niis used with 1sg and 1excl.pl, ki- is used for 1incl.pl as well as 2sg and 2pl. (6)

Qpyémbod mi-ge on / in ni-kné-mb-ih. yesterday 3-erg 1sg / 1excl.pl 1-see-ypst-3 ‘They saw me / us.excl yesterday.’7 b. Qpyémbod mi-ge mbi / ko / ki ki-kné-mb-ih. yesterday 3-erg 1excl.pl / 2sg / 2pl 2-see-ypst-3 ‘They saw us.incl/you.sg/you.pl yesterday.’

(7)

Qpyémbod in-a ko ki-kné-mb-ob. yesterday 1incl.pl-top 1sg 2-see-ypst-1pl ‘We saw you yesterday.’ b. Qpyémbod ko-ya in ni-kné-mb-e. yesterday 2sg-top 1incl.pl 2-see-ypst-2sg ‘You saw us yesterday.’

a.

a.

We might assume that, since the verbs code a nominative category by suffix, the prefixes are accusative. This hypothesis is, however, falsified when we examine a wider range of 6 The prefix ndo- can be used to apparently indicate a third person object, but without the obligatoriness found with inflection for first or second persons. This prefix has a number of additional functions, and should not be considered part of the paradigm with ni- and ki-. 7 Third person singular and plural are only rarely distinguished in Iha. To save space, all third person references will be translated as ‘they’/ ‘them’, although ‘he’, ‘she’, and ‘it’ are all equally appropriate.

410

mark donohue Table 17.1 The yesterday’s past inflection for hoqbów ‘fall’

1 12 2 3

sg

pl

ni-hoqbow-mb-on

ni-hoqbow-mb-ob ki-hoqbow-mb-ob ki-hoqbow-mb-i hoqbow-mb-ih

ki-hoqbow-mb-e hoqbow-mb-ih

data. The prefixes are also found with some monovalent verbs which encode uncontrolled predicates with patientive arguments. In (8) we can see that the single argument of hoqbów is indexed on the verb by both prefix and suffix. The full paradigm for this verb is shown in Table 17.1; note how the two first person plural categories found with free pronouns are distinguished in this kind of inflecting verb by the use of the different prefixes. Note, from (9), that an agentive, controlled verb such as ‘go’ does not allow the use of these prefixes (compare with (4b)). (8)

Qpyémbod ko-ya ki-hoqbów-mb-e. yesterday 2sg-top 2-fall-ypst-2sg ‘You fell over yesterday.’

(9)

∗ Qpyémbod

ko-ya Qpagqpag-na yesterday 2sg-top Fakfak-all ‘You went to Fakfak yesterday.’

ki-whó-mb-e. 2-go-ypst-2sg

Unlike the suffixes examined in Section 17.3.1, the prefixes seen here do not code a purely syntactic category, but a semantic one: patientive arguments are coded regardless of their syntactic role. While the suffixes code all and only S or A arguments, the prefixes code S or P arguments provided they are in some way patientive.8 For some monovalent verbs the only indication of the argument is by prefix. A number of verbs of experience have alternative codings, one which follows the pattern seen for hoqbów ‘fall’, and one transimpersonal construction with a ‘dummy’ third person coded by suffix. Both are shown in (10) with the verb qbáwo ‘sink’ (the difference, if any, between (9a) and (9b) is not clear, just as the availability of the different intransitive constructions, while influenced by verbal semantics, is at heart lexically determined). 8

Note that being ‘patientive’ is a necessary, but not sufficient, condition for a verb to show this prefixal agreement pattern. The verb kimi- ‘die’, perhaps the ultimate patientive verb, only shows suffixal agreement of the sort seen in Section 17.3.1. Thus we have kimiambe ‘You will die tomorrow’, and not ∗ ki -kimiambe with the prefixes described in this section.

verbal inflection in iha: a multiplicity of alignments (10) a. Mbi-ya dya lepalepa mur-mo-gana 1incl.pl-top boat canoe clf-that-ins ‘We sank in that canoe yesterday.’ b. Mbi-ya dya lepalepa mur-mo-gana 1incl.pl-top boat canoe clf-that-ins ‘We sank in that canoe yesterday.’ c. ∗ Mbi-ya qbáwo-mb-ob. 1incl.pl-top sink-ypst-1pl ‘We sank in that canoe yesterday.’

411

ki-qbáwo-mb-ob. 2-sink-ypst-1pl ki-qbáwo-mb-ih. 2-sink-ypst- ‘3’

In (11) we see the verb herered ‘(be) sick’, which only allows the transimpersonal construction. The undergoer of the predicate, marked with the topic case marker in (11), can show agreement on the verb by prefix only, and the suffixal agreement position is filled by the ‘third person’ suffixes. (11) a. Qpyémbod mbi-ya yesterday 1incl.pl-top ‘We were ill yesterday.’ b. ∗ Qpyémbod mbi-ya yesterday 1incl.pl-top ‘We were ill yesterday.’ c. ∗ Qpyémbod ko-ya yesterday 2sg-top ‘You were ill yesterday.’ d. Qpyémbod ko-ya yesterday 2sg-top ‘You were ill yesterday.’

ki-hereréd-mb-ih. 2-be.sick-ypst-‘3’ ki-hereréd-mb-ob. 2-be.sick-ypst-1pl ki-hereréd-mb-e. 2-be.sick-ypst-2sg ki-hereréd-mb-e. 2-be.sick-ypst-2sg

This ‘transimpersonal’ construction is found productively as an alternative to a regular transitive clause. This results in a construction vaguely remniscent of the ‘impersonal passive’ of English, or the Lingala passive construction (Dubinsky and Nzwanga 1994). In (11) we can see a typical bivalent clause, in which the NP representing the P is unmarked, but the verb shows agreement (by prefix) for this argument. From (12) we see that the object cannot be topic-marked, and yet in (13) the clause with no overt NP for A does allow a topic-marked ‘P’.9 This type of clause is frequently translated with a passive in Indonesian (the contact language used with informants who have formal education), but whether it represents a true passive (with quirky agreement) or not is a 9 The case marker -ya (and its allomorph -a) has a complex set of restrictions. It marks a participant of the clause as topical, including adjuncts, but of the core arguments only an S or A may be marked. The data in (13) thus suggests that on has a syntactic status different to that of the object seen in (14). Note that the fact that -ya can appear with non-arguments (principally time expressions) means that it cannot be simply called nominative case.

412

mark donohue

matter for further investigation. Note that the verb may not have the suffixal agreement that would be expected for most Ss (witness the ungrammaticality of (14)). (12)

a.

b.

(13)

(14)

Tambahár-ke on woman-erg 1sg ‘The woman saw me.’ ∗ Tambahár-ke

on-a woman-erg 1sg-top ‘The woman saw me.’

ni-knó-mb-ih. 1-see-ypst-3 ni-knó-mb-ih. 1-see-ypst-3

On-a ni-knó-mb-ih. 1sg-top 1-see-ypst-3 ‘I was seen.’ or ‘3SG saw me.’ ∗ On-a

ni-knó-mb-on. 1sg-top 1-see-ypst-sg1 ‘I was seen.’

17.3.3 Beneficiary Arguments and Verbal Agreement With beneficiaries we encounter another set of verbal suffixes, as seen in (15). The verb carries all normal inflection (compare (16)), and additionally carries a suffix marking the beneficiary. The paradigm for this category is verb small: there are only two members, -wen ‘1sg.ben’ and -win ‘1excl.pl.ben’, used for the first person plural inclusive, in (17). (15)

Mi-ge wri wy-wen-mb-ih. 3-erg house do-1SG.BEN-ypst-3 ‘They built a house for me.’

(16)

Mi-ge wri wy-mb-ih. 3-erg house do-ypst-3 ‘They built a house.’

(17)

Mi-ge wri wy-win- mb-ih. 3-erg house do-1excl.pl.ben-ypst-3 ‘They built a house for me.’

The incomplete paradigm is tolerated because the communicative function missing from the veb is taken up by a dedicated case marker, -wehad. This suffix is also available to mark beneficiary, shown in (18). The coding choice that we see by comparing of (15) and (18) does not exist for second or third persons.

verbal inflection in iha: a multiplicity of alignments (18)

413

Mi-ge wri on-wehad wy-mb-ih. 3-erg house 1sg-ben do-ypst-3 ‘They built a house for me.’

17.4 Suppletion

.............................................................................................................................................................................

We find significant suppletion with one verb, ‘give’, which simply changes the verb stem according to the person and number of the object (this is not unknown in other languages of New Guinea). (19)

a.

Qpót ków-mb-on. knife give.to.3-ypst-1sg ‘I gave her/him/them a knife.’

b.

Qpót kewé-mb-on. knife give.to.2sg-ypst-1sg ‘I gave you.sg a knife.’

c.

Qpót kiwi-mb-on. knife give.to.2pl-ypst-1sg ‘I gave you.pl a knife.’

d.

Qpót qpí-mb-ih. knife give.to.1incl.pl-ypst-3 ‘S/he gave us(.incl) a knife.’

e.

On qpót qpé-mb-ih. 1sg knife give.to.1-ypst-3 ‘S/he gave me a knife.’

f.

In qpót qpé-mb-ih. 1excl.pl knife give.to.1-ypst-3 ‘S/he gave us(.excl) a knife.’

These alternations in verb form only apply to the coding of a recipient in a (single) trivalent predicate, but the extended definition of alignment, referring also to the coding of themes and recipients with respect to Ps (Dryer 1986), is relevant. Since no known bivalent verbs in Iha show suppletion for the person or number of the P, the suppletion seen here is one way in which a P is treated in the same as a theme (by not being treated the same way as a recipient). Before we assume that Iha is a clear ‘indirect object’ language (following Dryer 1986), however, we should note that prefixal agreement, described earlier in Section 17.3.2, is also available to code the recipient, but never the theme, in a trivalent clause, as shown in (20). In terms of suppletion, then, a recipient is coded differently from a theme; in terms of pronominal prefixes, it is coded in the same way. The categorization of Iha trivalent constructions as showing

414

mark donohue Table 17.2 The verbs for ‘give’

1excl 1incl 2 3

sg

pl

qpe

qpe qpi kiwi kow

kewé kow

direct/indirect object, or primary/secondary object, coding strategies is unclear. Case marking does not disambiguate the two arguments, and so we cannot decide on this alignment question. (20)

a.

Mi-ge rámbutan on 3-erg rambutan 1sg ‘He gave me a rambutan.’

b.

Mi-ge on ko ki-kewé- da. 3-erg 1sg 2sg 2-give.to.2sg-tpst.3 ‘He gave me to you.’ (in a wedding exchange) Mi-ge 3-erg

ni-qpé-da. 1-give.to.1-tpst.3

c.

∗

on 1sg

ko 2sg

ni-kewé- da. 1-give.to.2sg-tpst.3

d.

∗ Mi-ge

on 1sg

ko 2sg

ni-qpe-da.10 1-give.to.1-tpst.3

3-erg

Suppletion is also found to code the number of some monovalent subjects; this has not been identified for many roots so far, but is obligatory with those verbs. Two such verbs, with preh / ngmbeh- ‘speak’ and tomot- / tidíh- ‘sleep’, are shown in (21) and (22): (21)

a.

On-a pret-qp-en. 1sg-top speak-prs-1sg ‘I am speaking.’ b. In-a ngmbeh-qp-eb. 1excl.pl-top speak.pl-prs-1pl ‘We are speaking.’ c.

∗ In-a

1excl.pl-top 10

pret-qp-eb. speak-prs-1pl

Grammatical with the meaning ‘He gave you to me.’

verbal inflection in iha: a multiplicity of alignments (22)

415

a.

On-a tomot-qp-en. 1sg-top sleep-prs-1sg ‘I am sleeping.’ b. In-a tidíh-qp-eb. 1excl.pl-top sleep.pl-prs-1pl ‘We are sleeping.’ ∗ c. In-a tomot-qp-eb. 1excl.pl-top speak-prs-1pl

17.5 One More Verbal Suffix

.............................................................................................................................................................................

In this section we examine complications to do with ‘conjugation class’ membership. Iha verbs are lexically split into three conjugation classes, without correlation with valency or agency. In some tenses the conjugation marker is not overt; in others, it is.11 In the examples presented so far there has been no overt marker of conjugation, since we have only seen examples in the present and yesterday’s past tenses. In Table 17.3 we can see different conjugation markers overtly appearing for both monovalent and bivalent verbs in the future, present, and today’s past (it is thus far from implausible to suggest that these ‘conjugation’ suffixes have some tense-related role).12 In (24) we can see the same tripartite choice for bivalent verbs; in the appropriate morphological environment, these conjugation suffixes must be present, with the choice of suffix lexically dictated by the verb. With some verbs, however, we find an alternation in the conjugational forms used. In (23) and (24) we can see that, in the yesterday’s past tense there is no overt realization of the conjugation marker, as expected. In (25) and (26), however, the conjugation marker is overt, and varies according to the number of the P: the ny-conjugation is found with a singular P, and the te-conjugation is found with a plural P. For verbs that display such a number-based conjugation change, the change is obligatory, as seen in the ungrammaticality of (27). Note, from Table 17.3, that there is nothing inherent in the ny-conjugation marker that is associated with plurality. Rather, some verbs (a lexically specified and semantically or phonological unpredicatable subset) utilize an alternation of conjugation types to indicate plurality.13 11 In some other dialects the conjugations are reportedly more regularly overt: purúny ombon, nórowg t ombon, for instance, in the southeastern Iha from Sipatnanam. 12 Note the allomorphy in the ny-conjugation marker, which loses the glide before a front vowel. The te-conjugation loses the vowel before other vowels, and so only occasionally surfaces. 13 Dukog ‘anim-small’ is commonly used with the lexicalized sense ‘child’, replacing wadin ‘child’ in most discourse. Possessed children are represented in a separate lexical item (namáda ‘my/our.excl child’, kamáda ‘your/our.incl child’, mandá, ‘third person’s child’)

416

mark donohue Table 17.3 Conjugation morphology with different verbs

Future Present Today’s past Yesterday’s past Remote past

Future Present Today’s past Yesterday’s past Remote past

yét ‘laugh’, 1sg Ø-conjugation

purú ‘fly’, 1sg ny-conjugation

nórowg ‘swim’, 1sg te-conjugation

yétanggen yéten yéten yétmbon yétnonggon

purúnyanggen purúnewen purúnen purúmbon purúnonggon

nórowgtanggen nórowgtewen nórowgten nórowgmbon nórowgnonggon

kmon ‘hear’, 1sg Ø-conjugation

qpréh ‘split wood’, 1sg ny-conjugation

qpag ‘hit’, 1sg te-conjugation

kmonanggen kmonen kmonen kmonmbon kmonnonggon

qpryéhnyanggen qpryéhnewen qpryéhnen qpryéhmbon qpryéhnonggon

qpagtanggen qpagtewen qpagten qpagmbon qpagnonggon

(23)

Mbyar du-mo du-kog dog anim-that anim-small ‘That dog bit one child.’

dyo-wo anim:one-one

(24)

Mbyar du-mo du-kog dog anim-that anim-small ‘That big dog bit two children.’

du-rig anim-two

(25)

Mbyar du-mo du-kog dog anim-that anim-small ‘That big dog bit one child.’

dyo-wo anim:one-one

(26)

Mbyar du-mo du-kog dog anim-that anim-small ‘That big dog bit two children.’

du-rig anim-two

wrig-t-a. bite-te [tpst]-3

du-rig anim-two

wrig-ny-a. bite-ny [tpst]-3

(27)

∗ Mbyar

du-mo du-kog dog anim-that anim-small ‘That big dog bit two children.’

wrig-mbih. bite-ypst-3

wrig-mbih. bite-ypst-3 wrig-ny-a. bite-ny [tpst]-3

This conjugation change, marking only number, interacts with the number-poor agreement prefixes described in Section 17.3.2 to maximally differentiate the person and number of objects. We have seen examples of third person differentiated in number on the verb by the change of conjugation marker, and in (28) and (29) we can see the utility of the conjugation change for local persons as well.

verbal inflection in iha: a multiplicity of alignments (28)

a.

Mbyar ni-wrig-ny-a. dog 1-bite-ny [tpst]-3 ‘The dog bit me.’

b.

Mbyar ni-wrig-t-a. dog 1-bite-te [tpst]-3 ‘The dog bit us(.excl).’

(29)

a.

Mbyar ki-wrig-ny-a. dog 2-bite-ny [tpst]-3 ‘The dog bit you.sg.’

b.

Mbyar ki-wrig-t-a. dog 2-bite-te [tpst]-3 ‘The dog bit you.pl/us.incl.’

417

This kind of conjugation class alternation is only found for bivalent verbs. Other verbs which are known to show this -ny/-te change are shown in (30). (30)

mbrang-conj ‘wash plates’, tongmbog-conj ‘split wood’, nggorog-conj ‘tie rope’, qprug-conj ‘stab/shoot’, ndo-qbri-conj ‘wash.tr’, phog-conj ‘blow at’.

Occasionally the alternation found is a -ny/-Ø alternation, not the common -ny/-te one described above. This alternative conjugation change is found with very few (so far only kangmbah ‘hit’ and ngmboho ‘throw’ have this pattern in our corpus). As before, all of the affected verbs are bivalent, and the change is always conditioned by the number of the P.

17.6 Conclusions

.............................................................................................................................................................................

We can summarize the different groupings that we have found in different parts of the verbs in Iha in Table 17.4. After examining these data, we might wonder what the ‘basic alignment type’ of Iha is. There is a clear nominative-accusative trend in the verbal morphology, but the presence of ‘non-agentive’ marking is also undeniable, suggesting (along with the lack of an unambiguous passive, and verbs such as ‘receive’ or ‘please’) that semantics plays

Table 17.4 Alignment categories found on the Iha verb P Verbal suffixes Stem suppletion Verbal prefixes Conjugations

S2

S1

A nominative (nominative?) non-agentive accusative

418

mark donohue

as much a role in Iha morphology as does syntax. While we have not explicitly discussed it, the examples shown make it clear that NP morphology for core arguments indicates pragmatic roles, rather than syntactic ones, and indicates that a larger split in morphological orientation can be found in the Iha clause. We have seen that verbal agreement in Iha expounds a range of different coding choices, with nominative, accusative, and patientive categories all relevant to the description of verbal alignment. It is certainly not unusual for more than one ‘alignment system’ (for instance, ergative-absolutive or nominative-accusative) to be relevant in a single language. In (1) and (2) we saw data from Nepali, a language with ergative case marking on NPs but nominative agreement on Vs, and other examples of this would not be hard to come by, in New Guinea (Donohue 2005) or elsewhere. Equally, it is quite possible for more than one system to be coded on a predicate. In (written) French, verbs mark the number of the nominative argument on the verb, but the gender of the non-active argument is marked on the participle (when the object is pronominal, for bivalent verbs), as shown in (30) and (31). (30) French a.

Il l’a mangé. 3sg.m 3sg-have eaten ‘He ate itmasculine .’

b.

Il l’a 3sg.m 3sg-have ‘He ate itfeminine .’

mange-e. eaten-f

b.

Elle est parti-e. 3sg.f is left-f ‘She has gonefeminine .’

(31) French a.

Il est parti. 3sg.m is left ‘He has gone.’

The Iha data are unusual in the presentation of four different alignment categories which are all relevant to the assignment of agreement, in sometimes overlapping ways. This is not impossible (Bickel 2004 discusses a language with many non-morphological patterns of alignment), but it is a striking rebuff to the notion that even in a subdomain of a language it is possible to adequately label ‘alignment’ with a single word.

chapter 18 ........................................................................................................

INFLECTION IN PULAAR ........................................................................................................

fiona mc laughlin

18.1 Introduction

.............................................................................................................................................................................

Pulaar is the westernmost dialect of Fula, a language noted for having one of the largest gender systems found in natural language and for the way in which it expresses much of its inflectional morphology through consonant mutation. Fula is an Atlantic language of the Niger-Congo phylum whose speakers are dispersed across a 5,000 or so kilometre stretch of the West African Sahel, aptly dubbed the Fula archipelago by Boutrais (1994). The dispersion of Fula speakers from Senegal and Mauritania in the west to Chad and Sudan in the east is a consequence of their historical mobility as pastoralists, but there are also many sedentary Fula-speaking groups. By the most conservative estimates there are at least 12 to 15 million speakers of the language, but Fula is a minority language in every country in which it is spoken, with the possible exception of Guinea, and as a result most speakers are bilingual or multilingual. The westernmost dialects of Fula more often than not go by a p-initial variant of the name for the language: Pulaar (Senegal), Pular (Guinea), and Peul in French (from Wolof Pël); the central and eastern dialects on the other hand are f-initial: Fulfulde and Fulani, the latter being from Hausa. The p/f alternation illustrates the phenomenon of initial consonant mutation, a salient aspect of Fula inflectional morphology discussed in some detail in this chapter. The overall architecture of Fula grammar is much the same across dialects, but there are significant phonological, morphological, and syntactic differences between them, as well as lexical differences, many of which have their origin in the diverse languages with which Fula has been in contact. Fula speakers constitute many different ethnic groups across the Sahel, including but not limited to transhumant pastoralists or Fulɓe, the semi-nomadic Woɗaaɓe of Niger, and the sedentary and long islamized Haalpulaar’en (often referred to in the literature as Tukuloor or Toucouleur) who live on both banks of the Senegal River in southern Mauritania and northern Senegal. The dialect discussed in this chapter is

420

fiona mc laughlin

Fuutankoore Pulaar, the dialect of the Fuuta Tooro region of northern Senegal and southern Mauritania whose speakers call themselves Haalpulaar’en. Pulaar is a conservative Fula dialect and is spoken in or near the region in which the language originated. Fula’s closest sister language, Seereer-Siin, and the more distant yet structurally similar Wolof, both of which are spoken in Senegal, will be discussed briefly in Section 18.5 to provide a comparative perspective.

18.2 Grammatical Context

.............................................................................................................................................................................

Typologically, Pulaar is a relatively synthetic and somewhat fusional language that exhibits many of the grammatical characteristics typical of the Niger-Congo languages in general and the Atlantic languages in particular.1 It has a basic SVO word order, a robust gender or noun class system, a small number of underived adjectives, and a set of derivational verbal extensions. Nouns, their determiners, and concordant adjectives, as well as third person pronouns and anaphora, are inflected for noun class, but in contradistinction to the Bantu languages, subject and object class markers do not appear on the verb. Pulaar exhibits no inflectional marking of case on nouns, but distinguishes between subject and object pronouns, although not between accusative and dative pronouns. Pulaar has a very small number of underived adjectives, and a larger number of adjectives derived from nouns and from stative verbs of the kind wulde ‘to be hot’, hoyde ‘to be light’, etc. Degree is handled by a periphrastic construction that utilizes the verb ɓurde ‘to exceed or surpass’, in conjunction with an adjectival verb, thus it is not an inflectional category in Pulaar. There are also some instances of inherently comparative or superlative adjectival verbs such as heccude ‘to be older than’ and faɗɗude ‘to be more beautiful than’. The Pulaar pronominal system exhibits an inclusive/exclusive distinction in the first person plural. In typical Niger-Congo fashion there is no distinction based on gender in the pronominal system; third person anaphora, however, are inflected for the noun class of their referent, giving rise to a striking multiplicity of forms. Pulaar verbs are inflected for voice (active, middle, and passive), aspect (perfective and imperfective), and polarity. In addition, there is a suite of morphosyntactic properties that go to make up a category that has been referred to in the literature on Fula as ‘relative tense’, and which I use interchangeably with the more abstract term, R-construction, in this chapter. Included in this category are narrative clauses, relative clauses, wh-phrases, phrases introduced by the temporal marker nde ‘when’, and clauses in which the salience 1 It is a matter of some debate as to whether Atlantic (formerly West Atlantic) is a coherent genetic subgroup of Niger-Congo. There appears to be no shared innovation between the Atlantic languages, and rates of cognacy are fairly low, so they may well simply be a typological grouping. There is enough similarity between the Senegal languages of Northern Atlantic, however, to maintain a close genetic relationship between those languages, which include Fula, Seereer, and Wolof.

inflection in pulaar

421

of some constituent has been increased and therefore promoted on the information hierarchy. The habitual use of quotation marks around the term ‘relative tense’ (Arnott 1970; McIntosh 1984; Gottschligg 2006; etc.) implies less than full satisfaction with the term, and further elucidation of the distinction between relative and general verbal forms awaits a better understanding of information and discourse structure in Pulaar. What is clear, however, is that information within the Pulaar clause is organized by a combination of syntactic and pragmatic considerations that are highly discourse sensitive. In addition to clefts and pseudo-clefts which focus constitutents, there is a focus particle, ko, which marks identificational focus in specificational, but not predicational, clauses (Cover 2009), and which also has a copular function. Verbs are not inflected for tense, but a preterite or aorist marker /-no(o)/ is found across Fula dialects. McIntosh (1984: 76) calls it an ‘anteriority marker’ and says that ‘[e]ssentially, it sets the action depicted by the verbal complex one step back in time, using the moment of utterance or a time determined in the discourse as its point of reference’. Gottschligg (2006: 147) claims that the anteriority marker is not part of the inflectional system of Fula and that it can attach to non-verbal forms in some dialects, thereby consolidating its status as a clitic.2 Finally, verbs in Pulaar are not inflected for person, and are only minimally inflected for number through stem-initial consonant mutation.

18.3 Features, Forms, and Paradigms

.............................................................................................................................................................................

18.3.1 Phonological and Morphological Overview Pulaar is primarily a suffixing language as illustrated in the examples in (1) and (2). There is no overt prefixation in the language, although consonant mutation can be analysed as an abstract kind of prefixation as we shall see in Section 18.3.1.1. Reduplication is not very productive, and in any case is limited to derivational processes.3 In (2), the noun class marker occurs as a suffix to a nominal stem in contradistinction to the Bantu languages where class markers occur as prefixes.4 (1) haal-t-ii-no speak.sg-iter-pfv-ant ‘spoke again’ 2

This issue merits further probing, but it is perhaps relevant to note that Zribi-Hertz and Diagne (2002: 831) also raise questions about the puzzling distribution of the (possibly cognate) past marker /-oon/ in Wolof, although they do not go so far as to claim that it is a clitic. 3 Although restricted, reduplication in Pulaar is of significant theoretical interest because of its interaction with consonant mutation. See Mc Laughlin (2005) for a discussion of this issue in comparative Northern Atlantic perspective. 4 See Greenberg (1977) for a historical perspective on how eroded noun class prefixes were ‘renewed’ as suffixes in Fula.

422

fiona mc laughlin

(2) ngil-ngu earthworm-clngu ‘earthworm’ Besides suffixation, much of the inflectional morphology of Pulaar involves stem alternations such as those found in the singular/plural pairings of nouns5 in (3) and verbs in (4). (3) Singular class hoo-re ngaa-ri ree-du wow-ru gawl-o dim-o (4) Singular hirtyahrokk-

Plural class koʔ-e 6 gaʔ-i dee-di boɓ-i ʔawl-(u)ɓe rim-ɓe

Plural & R kirtnjahndokk-

‘head’ ‘bull’ ‘belly’ ‘mortar’ ‘griot’ ‘noble’

‘dine’ ‘go’ ‘give’

These patterns of alternation, referred to in the literature as consonant mutation, are found widely in the Atlantic languages. In Pulaar they are generally regular and extend even to the morphological integration of loanwords in the language.7 Consonant mutation has been the subject of intense theoretical inquiry, much of which revolves around the question of whether it reflects an Item-and-Process or an Item-and-Arrangement type of mapping between phonological form and morphosyntax (Hockett 1954), and there is as yet no consensus. It is possible to treat mutating stems as lexical lists of allomorphs as Green (2006) has done for similar phenomena in the Celtic languages where mutation is not as systematic, but an alternative autosegmental treatment can also account successfully for the systematicity of Pulaar mutations (Lieber 1984, 1987; Wolf 2007; etc.). In the autosegmental approach adopted here, stem-initial consonant mutation is understood as the result of the prefixation of a morpheme whose exponence is smaller than a segment, namely a floating feature or autosegment, which is realized on the stem-initial consonant. There are two and only two such featural exponences in Pulaar, [−continuant] and [+nasal]. Morphemes with both exponences play a role in nominal inflection (noun class), while 5 Number and noun class are interdependent, so although I have listed the singular and plural forms of these nouns, it is actually their noun class that determines the initial consonant. 6 An intervocalic glottal stop is often pronounced as a palatal glide before a front vowel. 7 Arnott (1970: 109) also states that ‘assimilated’ loanwords in the Gombe (Nigeria) dialect of Fula exhibit consonant mutation, resulting in what, from a pre-autosegmental perspective, he considers to be backformations. Breedveld (1995: 69), on the other hand, reports that loanwords in the Malian dialect of Maasinankoore tend not to undergo consonant mutation.

inflection in pulaar

423

Table 18.1 The consonants of Pulaar Voiceless plosives Voiced plosives Implosives Fricatives Nasals Prenasalized stops Liquids

p b ɓ f m mb

Glides

w

t d ɗ s n nd r l

c j ʄ

k g

ñ nj

ŋ ng

y

w

h

Table 18.2 Pulaar gradation setsa Labial Continuant Stop Nasal

w b mb

Coronal f p p

r d nd

s c c

Dorsal y j nj

h k k

y g ng

w g ng

ʔ g ng

a Voiceless prenasalized stops are not permitted in Pulaar;

voiceless consonants occur in stop form in the nasal grade.

verbal inflection makes use of only the latter (plurality and R). If we accept this view, then Pulaar can be said to have prefixes, albeit abstract ones with minimal exponence. The fact that they are the only prefixes in the language is compatible with the historical and comparative facts since noun class markers are the only prefixes in many Niger-Congo languages. The following discussion, which necessarily delves into the phonology of Pulaar, provides the necessary background information for understanding the abstract nature of inflectional exponence evidenced in Pulaar consonant mutation.

18.3.1.1 The phonology of consonant mutation Of the set of Pulaar consonants, given in Table 18.1, all but the voiceless stop /t/, the implosive consonants /ɓ/, /ɗ/, and /ʄ/, the plain nasals /m/, /n/, /ñ/ and /ŋ/, and liquid /l/ participate in the mutation system.8 Consonants may alternate between three homorganic grades, a continuant, a stop, and a nasal (prenasalized stop). There are nine such gradation sets in Pulaar (Table 18.2). 8 Here I follow the conventional Senegalese Roman script orthography for Pulaar with one exception: I use the IPA symbol [ʄ] for the palatal implosive stop which is conventionally written as a hooked y.

424

fiona mc laughlin

The prefixation of [−continuant] or [nasal] to a continuant-initial stem such as /wur-/ ‘village’ results in the outputs gur- and ngur-, respectively. Prefixation of [−continuant] to a non-continuant-initial stem such as /deft-/ ‘book’, or [nasal] to a prenasalized stop-initial stem such as /ngool-/ ‘cobra’, on the other hand, is vacuous and results in no change. Prenasalized stop-initial stems, then, do not exhibit consonant mutation; stop-initial stems show a two-way alternation; and continuant-initial stems may exhibit all three grades, as illustrated in the following examples. (5) Gradation with different stem types Base [−continuant] Continuant initial worgorStop initial juljulNasal initial ngool- ngool-

[nasal] ngornjulngool-

‘man’ ‘trader’ ‘cobra’

The behaviour of French loanwords provides further evidence for such an approach to consonant mutation. Given the appropriate morphological environment, continuantinitial stems such as /raŋ-/ from French rang ‘queue’ can become daŋ- or ndaŋ-, while stop-initial stems such as /diskit-/ from French discuter ‘to argue’ can become ndiskit- but never riskit-, because while there are noun class morphemes whose exponence is [−continuant] and [nasal], there are no morphemes whose exponence is [+continuant]. Finally, because Pulaar does not allow onsetless syllables, a glottal stop is epenthesized to vowel-initial French borrowings and the glottal stop then undergoes mutation in the right morphological environment; thus the French verb ourler ‘to hem’ is borrowed as the Pulaar stem /ʔurl-/ and gives rise to the related forms gurland ngurl- (Mc Laughlin 2011). The patterns of consonant mutation, then, are quite regular, but there are a number of native stems like hoggo ‘beak’ that do not undergo mutation and can therefore be considered lexical exceptions: (6)

Singular Plural Diminutive plural

hoggo hoggooji ɗı hoggoyon kon

ngo ∗ koggooji

ɗı kon

∗ koggoyon

18.3.2 Features The inflectional features of Pulaar which will be discussed in this section, include gender, which manifests itself as noun class, case as seen in the pronominal system, number, the trio voice, aspect, and polarity, all of which are expressed cumulatively on the verb, along with salience (traditionally called focus or emphasis) and the related abstract category R.9 Pulaar verbs are not inflected for person. 9 Recall from Section 18.2 that R represents the complex of constructions that constitute what have been called the ‘relative tenses’.

inflection in pulaar

425

18.3.2.1 Noun class Except for the fact that they typically consist of a much larger number of classes, and gender does not play any role in their semantics (the words for man (gorko) and woman (debbo), for example, are both in the human singular o-class), Niger-Congo noun class systems are similar in most respects to the gender systems found in Indo-European and Afro-Asiatic languages in that they trigger agreement throughout the noun phrase; noun class and gender are consequently conflated in the literature as a single inflectional feature. Here I retain the term noun class solely as an Africanist convention without any theoretical implications. In the Pulaar noun class system there are twenty-one classes, conventionally referred to by their corresponding definite article. Noun class is marked on the inflected noun or adjective by a class suffix and its concomitant stem-initial consonant grade. With the exception of a very small number of native nouns and a larger number of loanwords that fall into the o-class (the only class which does not require a noun class suffix) nouns are obligatorily inflected for class and do not occur as bare stems.10 The classes and their definite articles are listed in Table 18.3, along with the initial consonant grade they condition. The class suffix and the independent definite article are quite similar in phonological form, but the article is nonetheless a distinct word. Historically, the definite article is the source of the suffix (Greenberg 1977), and the closer morphological relationship between suffix and noun stem has resulted in certain phonological changes at the morpheme boundary so that class suffixes also exhibit allomorphy in the form of initial consonant mutation, as illustrated in the examples in (7) where morpheme boundaries are indicated between stem and suffix.11 In general, the grade of the suffix-initial consonant (or its absence) is a lexical property of the stem, although the derivational status of the stem, and sometimes its phonological shape, may also play a role in determining the grade. (7) ngol-class suffixes kuɗ-ol ɗere-wol jaah-gol gon-ngol

‘blade of grass’ ‘sheet of paper’ ‘departure’ ‘tear’

no initial consonant continuant initial stop initial prenasalized stop initial

Finally, allomorphy in Pulaar nouns can also be seen sometimes in the stem-final consonant, as in the singular (o-class) and plural (ɓe-class) forms of the word for ‘woman’: 10 At a slightly more abstract level we could, of course, posit a Ø-suffix for such nouns. Gottschligg (2006: 146) gives some examples of suffixless forms of nouns in Pular, the Fuuta Jaloo (Guinea) dialect of Fula, where they have a generic reading, e.g. gerto ‘chicken’ (as meat) vs. gerto-gal ‘chicken’ (the bird). 11 In what follows, morpheme boundaries between nominal or adjectival stem and suffix will be marked (e.g. kuɗ-ol) only when relevant to the discussion at hand. Boundaries will not be marked in citation forms (e.g. kuɗol). Certain phonological processes such as assimilation and gemination that take place at the boundary between noun or adjective stem and class suffix can render the boundary somewhat difficult to delineate. In the examples in (7), however, they are quite clear.

426

fiona mc laughlin

debb-o ‘woman’, rew-ɓe ‘women’. Such alternation of the stem-final consonant is for the most part phonologically conditioned, or was so historically and has now become lexicalized. The grade of the stem-initial consonant, on the other hand, is conditioned by noun class. Of the twenty-one classes, six condition the continuant grade, eight the stop grade, and seven the nasal grade, thus the distribution of grades is more or less even across classes, with each grade appearing in seven (±1) classes. There is a semantic core to the system but it is far from being transparent. Semantic categories include natural, physical, and metaphorically extended properties. For example, the ɗam class includes liquids (eg: ndiyam ‘water’, kosam ‘milk’, ʄiiʄam ‘blood’); the ngol class includes long, extended objects or concepts (e.g. caɗngol ‘river’, ɓoggol ‘rope’, lefol ‘woven strip’, koyɗol ‘dream’, jimol ‘song’, leñol ‘lineage’, timtimol ‘rainbow’); the ki class includes trees (e.g. lekki ‘tree’, ɓokki ‘baobab’, darkaseewi ‘cashew nut tree’). The small nge class includes bovines (e.g. nagge ‘cow’, wiige ‘heifer’), which are culturally highly salient for the transhumant Fulɓe, but this class also includes naange ‘sun’, which is related to bovines by the myth-and-belief principle (Lakoff 1986), and five other semantically disparate items.12 Although these and other classes may contain a semantic core, they almost always include a great number of nouns that fall outside those categories, rendering the semantics somewhat opaque as is typical of noun class systems in general.13 The ɗum-class is a neutral or indeterminate class used for unspecified items as seen in the following examples: 14 (8)

Ko ɗum foc 3sgɗum ‘What is that?’

(9)

oolum ɗum yellowɗum detɗum ‘That which is yellow’

woni? be.sg-pfv.r

Morphological considerations can also play a role in noun class assignment. The o-class, for example, is a human singular class (eg: neɗɗo ‘person’, debbo ‘woman’), and almost all human singular nouns fall into it unless they are assigned to the diminutive or augmentative classes,15 but it is also the default class for loanwords because it 12

Because it is such a small class, consisting of a total of only six nouns (of which some, like liige ‘cotton field’, have a low frequency) in addition to those that designate bovines, knowledge of all members of the class is considered a playful diagnostic for identifying erudite speakers of the language (Abdoulaye Kane p.c.). 13 For a thorough and detailed discussion of the semantics of the noun class system of Maasinankoore, a Malian dialect of Fula, see Breedveld (1995). 14 See also Section 18.3.2.1.3 on how ɗum-class possessive pronouns can occur coreferentially with overt nominal antecedents in a different class. 15 Mboomri ‘young girl, virgin’ is one of the few human nouns that does not fall into the human o-class; it is assigned rather to the ndi-class.

inflection in pulaar

427

Table 18.3 Noun classes, grades, and core semantic content Singular

Plural

Class Grade

Content

o ndu ngu nge ngol nde ndi ngo ki ɗam ngal ba ka ko ɗum ngel kal

humans, loanwords ɓe hollow, round, inflated ɗi ɗe sun, bovines kon long, extended

stop continuant continuant continuant stop continuant nasal continuant nasal nasal stop nasal nasal continuant nasal stop stop

Class Grade

Content

continuant plural, human stop stop nasal diminutive

curved up at edges trees, straight upright liquids

neutral, indeterminate diminutive diminutive (mass-nouns)

does not require an overt class marker (e.g.: njambala ‘giraffe’ (